Update README section on powered research

fix(whisper/tokenizer): prevent IndexError from crashing multilingual…

.dockerignore

@@ -0,0 +1,13 @@
+.git
+.github
+.venv
+__pycache__
+*.pyc
+.pytest_cache
+.mypy_cache
+.ruff_cache
+.cache
+.tmp
+.secrets
+dist
+build

.github/workflows/publish-docker.yml

@@ -0,0 +1,61 @@
+name: Publish Docker Images
+
+on:
+  push:
+    tags:
+      - "v*"
+  workflow_dispatch:
+    inputs:
+      tag:
+        description: "Image tag to publish (without image suffix)"
+        required: true
+        type: string
+
+permissions:
+  contents: read
+  packages: write
+
+jobs:
+  docker:
+    runs-on: ubuntu-latest
+    env:
+      IMAGE_TAG: ${{ github.event_name == 'workflow_dispatch' && github.event.inputs.tag || github.ref_name }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - image_suffix: cpu-diarization-sortformer
+            dockerfile: Dockerfile.cpu
+            extras: cpu,diarization-sortformer
+          - image_suffix: cu129-diarization-sortformer
+            dockerfile: Dockerfile
+            extras: cu129,diarization-sortformer
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set lowercase owner
+        id: owner
+        run: echo "value=${GITHUB_REPOSITORY_OWNER,,}" >> "${GITHUB_OUTPUT}"
+
+      - name: Login to GHCR
+        uses: docker/login-action@v3
+        with:
+          registry: ghcr.io
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Setup Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Build and push image
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: ./${{ matrix.dockerfile }}
+          push: true
+          build-args: |
+            EXTRAS=${{ matrix.extras }}
+          tags: |
+            ghcr.io/${{ steps.owner.outputs.value }}/whisperlivekit:${{ env.IMAGE_TAG }}-${{ matrix.image_suffix }}
+            ghcr.io/${{ steps.owner.outputs.value }}/whisperlivekit:latest-${{ matrix.image_suffix }}

.gitignore

@@ -119,9 +119,11 @@ run_*.sh
 *.pt 
 
 # Debug & testing
-test_*.py
+/test_*.py
+!test_backend_offline.py
 launch.json
 .DS_Store
-test/*
+/test/
+!tests/
 nllb-200-distilled-600M-ctranslate2/*
 *.mp3

BENCHMARK.md

@@ -0,0 +1,205 @@
+# WhisperLiveKit Benchmark Report
+
+Benchmark comparing all supported ASR backends, streaming policies, and model sizes on Apple Silicon.
+All tests run through the full AudioProcessor pipeline (same code path as production WebSocket).
+
+## Test Environment
+
+| Property | Value |
+|----------|-------|
+| Hardware | Apple M4, 32 GB RAM |
+| OS | macOS 25.3.0 (arm64) |
+| Python | 3.13 |
+| faster-whisper | 1.2.1 |
+| mlx-whisper | installed (via mlx) |
+| Voxtral MLX | native MLX backend |
+| Voxtral (HF) | transformers-based |
+| VAC (Silero VAD) | enabled unless noted |
+| Chunk size | 100 ms |
+| Pacing | no-realtime (as fast as possible) |
+
+## Audio Test Files
+
+| File | Duration | Language | Speakers | Description |
+|------|----------|----------|----------|-------------|
+| `00_00_07_english_1_speaker.wav` | 7.2 s | English | 1 | Short dictation with pauses |
+| `00_00_16_french_1_speaker.wav` | 16.3 s | French | 1 | French speech with intentional silence gaps |
+| `00_00_30_english_3_speakers.wav` | 30.0 s | English | 3 | Multi-speaker conversation |
+
+Ground truth transcripts (`.transcript.json`) with per-word timestamps are hand-verified.
+
+---
+
+## Results
+
+### English -- Short (7.2 s, 1 speaker)
+
+| Backend | Policy | Model | RTF | WER | Timestamp MAE |
+|---------|--------|-------|-----|-----|---------------|
+| faster-whisper | LocalAgreement | base | 0.20x | 21.1% | 0.080 s |
+| faster-whisper | SimulStreaming | base | 0.14x | 0.0% | 0.239 s |
+| faster-whisper | LocalAgreement | small | 0.59x | 21.1% | 0.089 s |
+| faster-whisper | SimulStreaming | small | 0.39x | 0.0% | 0.221 s |
+| mlx-whisper | LocalAgreement | base | 0.05x | 21.1% | 0.080 s |
+| mlx-whisper | SimulStreaming | base | 0.14x | 10.5% | 0.245 s |
+| mlx-whisper | LocalAgreement | small | 0.16x | 21.1% | 0.089 s |
+| mlx-whisper | SimulStreaming | small | 0.20x | 10.5% | 0.226 s |
+| voxtral-mlx | voxtral | 4B | 0.32x | 0.0% | 0.254 s |
+| voxtral (HF) | voxtral | 4B | 1.29x | 0.0% | 1.876 s |
+
+### English -- Multi-speaker (30.0 s, 3 speakers)
+
+| Backend | Policy | Model | RTF | WER | Timestamp MAE |
+|---------|--------|-------|-----|-----|---------------|
+| faster-whisper | LocalAgreement | base | 0.24x | 44.7% | 0.235 s |
+| faster-whisper | SimulStreaming | base | 0.10x | 5.3% | 0.398 s |
+| faster-whisper | LocalAgreement | small | 0.59x | 25.0% | 0.226 s |
+| faster-whisper | SimulStreaming | small | 0.26x | 5.3% | 0.387 s |
+| mlx-whisper | LocalAgreement | base | 0.06x | 23.7% | 0.237 s |
+| mlx-whisper | SimulStreaming | base | 0.11x | 5.3% | 0.395 s |
+| mlx-whisper | LocalAgreement | small | 0.13x | 25.0% | 0.226 s |
+| mlx-whisper | SimulStreaming | small | 0.20x | 5.3% | 0.394 s |
+| voxtral-mlx | voxtral | 4B | 0.31x | 9.2% | 0.176 s |
+| voxtral (HF) | voxtral | 4B | 1.00x | 32.9% | 1.034 s |
+
+<p align="center">
+<img src="benchmark_chart.png" alt="Benchmark comparison on 30s English" width="800">
+</p>
+
+<p align="center">
+<img src="benchmark_scatter.png" alt="Speed vs Accuracy tradeoff" width="700">
+</p>
+
+### French (16.3 s, 1 speaker, `--language fr`)
+
+| Backend | Policy | Model | RTF | WER | Timestamp MAE |
+|---------|--------|-------|-----|-----|---------------|
+| faster-whisper | LocalAgreement | base | 0.22x | 25.7% | 3.460 s |
+| faster-whisper | SimulStreaming | base | 0.10x | 31.4% | 3.660 s |
+| faster-whisper | LocalAgreement | small | 0.76x | 42.9% | 0.051 s |
+| faster-whisper | SimulStreaming | small | 0.29x | 25.7% | 0.219 s |
+| mlx-whisper | LocalAgreement | base | 0.09x | ~45%* | ~5.0 s* |
+| mlx-whisper | SimulStreaming | base | 0.09x | 40.0% | 3.540 s |
+| mlx-whisper | LocalAgreement | small | 0.14x | 25.7% | 0.083 s |
+| mlx-whisper | SimulStreaming | small | 0.17x | 31.4% | 0.203 s |
+| voxtral-mlx | voxtral | 4B | 0.18x | 37.1% | 3.422 s |
+| voxtral (HF) | voxtral | 4B | 0.63x | 28.6% | 4.040 s |
+
+\* mlx-whisper + LocalAgreement + base is unstable on this French file (WER fluctuates 34-1037% across runs due to hallucination loops). The `small` model does not have this problem.
+
+**Timestamp note:** The base model produces very high timestamp MAE (3.4-3.7s) on this French file because it misaligns words around the silence gaps. The small model handles this much better (0.05-0.22s MAE). Voxtral also drifts on the silence gaps.
+
+---
+
+## Model Size Comparison (base vs small)
+
+| | base | small | Observation |
+|--|------|-------|-------------|
+| **RTF** | 0.05-0.24x | 0.13-0.76x | small is 2-3x slower |
+| **English WER (SS)** | 0-5.3% | 0-5.3% | No improvement: SimulStreaming already saturates on base |
+| **English WER (LA)** | 21-44.7% | 21-25% | small reduces LA errors on longer audio |
+| **French WER** | 25-40% | 25-43% | Mixed: depends on backend/policy combo |
+| **French timestamps** | 3.4-5.0s MAE | 0.05-0.22s MAE | small is dramatically better for French timestamps |
+
+In short: **base + SimulStreaming** gives the best speed/accuracy tradeoff for English. The small model only helps if you need LocalAgreement (for subtitle-grade timestamps) or non-English languages.
+
+---
+
+## Key Findings
+
+### Speed (RTF = processing time / audio duration, lower is better)
+
+1. **mlx-whisper + LocalAgreement + base** is the fastest combo on Apple Silicon: 0.05-0.06x RTF on English. 30 seconds of audio in under 2 seconds.
+2. For **faster-whisper**, SimulStreaming is faster than LocalAgreement. For **mlx-whisper**, it is the opposite: LocalAgreement (0.05-0.06x) outperforms SimulStreaming (0.11-0.14x) on speed.
+3. **voxtral-mlx** runs at 0.18-0.32x RTF -- 3-5x slower than mlx-whisper base, but well within real-time.
+4. **voxtral (HF transformers)** hits 1.0-1.3x RTF. At the real-time boundary on Apple Silicon. Use the MLX variant instead.
+5. The **small** model is 2-3x slower than base across all backends.
+
+### Accuracy (WER = Word Error Rate, lower is better)
+
+1. **SimulStreaming** gives dramatically lower WER than LocalAgreement on the whisper backends. On the 30s English file: 5.3% vs 23-44%.
+2. **voxtral-mlx** hits 0% on short English and 9.2% on multi-speaker. It auto-detects language natively. Whisper also supports `--language auto`, but tends to bias towards English on short segments.
+3. **LocalAgreement** tends to repeat the last sentence at end-of-stream (a known LCP artifact), inflating WER. This is visible in the 21% WER on the 7s file -- the same 4 extra words appear in every LA run.
+4. On **French** with the correct `--language fr`, whisper base achieves 25-40% WER -- comparable to Voxtral's 28-37%. The small model does not consistently improve French WER.
+
+### Timestamps (MAE = Mean Absolute Error on word start times)
+
+1. **LocalAgreement** gives the best timestamps on English (0.08-0.09s MAE).
+2. **SimulStreaming** is less precise (0.22-0.40s MAE) but good enough for most applications.
+3. On French with silence gaps, **base model timestamps are unreliable** (3.4-5s MAE). The **small model fixes this** (0.05-0.22s MAE). This is the strongest argument for using `small` over `base`.
+4. **voxtral-mlx** has good timestamps on English (0.18-0.25s MAE) but drifts on audio with long silence gaps (3.4s MAE on the French file).
+
+### VAC (Voice Activity Classification) Impact
+
+| Backend | Policy | VAC | 7s English WER | 30s English WER |
+|---------|--------|-----|----------------|-----------------|
+| faster-whisper | LocalAgreement | on | 21.1% | 44.7% |
+| faster-whisper | LocalAgreement | off | 100.0% | 100.0% |
+| voxtral-mlx | voxtral | on | 0.0% | 9.2% |
+| voxtral-mlx | voxtral | off | 0.0% | 9.2% |
+
+- **Whisper backends need VAC** to work in streaming mode. Without it the buffer logic breaks down and you get empty or garbage output.
+- **Voxtral is unaffected by VAC** since it handles its own internal chunking. Identical results with or without. VAC still saves compute on silent segments.
+
+---
+
+## Recommendations
+
+| Use Case | Backend | Policy | Model | Notes |
+|----------|---------|--------|-------|-------|
+| Fastest English (Apple Silicon) | mlx-whisper | SimulStreaming | base | 0.11x RTF, 5.3% WER |
+| Fastest English (Linux/GPU) | faster-whisper | SimulStreaming | base | 0.10x RTF, 5.3% WER |
+| Best accuracy, English | faster-whisper | SimulStreaming | small | 0.26x RTF, 5.3% WER, still fast |
+| Multilingual / auto-detect | voxtral-mlx | voxtral | 4B | 100+ languages, 0.18-0.32x RTF |
+| Best timestamps | any | LocalAgreement | small | 0.05-0.09s MAE, good for subtitles |
+| Low memory / embedded | mlx-whisper | SimulStreaming | base | Smallest footprint, fastest response |
+
+---
+
+## Caveats
+
+- **3 test files, ~53 seconds total.** Results give relative rankings between backends but should not be taken as definitive WER numbers. Run on your own data for production decisions.
+- **RTF varies between runs** (up to +/-30%) depending on thermal state, background processes, and model caching. The numbers above are single sequential runs on a warm machine.
+- **Only base and small tested.** Medium and large-v3 would likely improve WER at the cost of higher RTF. We did not test them here because they are slow on Apple Silicon without GPU.
+
+---
+
+## Reproducing These Benchmarks
+
+```bash
+# Install test dependencies
+pip install -e ".[test]"
+
+# Single backend test
+python test_backend_offline.py --backend faster-whisper --policy simulstreaming --model base --no-realtime
+
+# With a specific language
+python test_backend_offline.py --backend mlx-whisper --policy simulstreaming --model small --lan fr --no-realtime
+
+# Multi-backend auto-detect benchmark
+python test_backend_offline.py --benchmark --no-realtime
+
+# Export to JSON
+python test_backend_offline.py --benchmark --no-realtime --json results.json
+
+# Test with your own audio
+python test_backend_offline.py --backend voxtral-mlx --audio your_file.wav --no-realtime
+```
+
+The benchmark harness computes WER and timestamp accuracy automatically when ground truth
+`.transcript.json` files exist alongside the audio files. See `audio_tests/` for the format.
+
+---
+
+## Help Us Benchmark on More Hardware
+
+These results are from a single Apple M4 machine. We'd love to see numbers from other setups: Linux with CUDA GPUs, older Macs, different CPU architectures, cloud instances, etc.
+
+If you run the benchmark on your hardware, please open an issue or PR with your results and we will add them here. The more data points we have, the better the recommendations get.
+
+What we are especially interested in:
+- **NVIDIA GPUs** (RTX 3090, 4090, A100, T4, etc.) with faster-whisper
+- **Older Apple Silicon** (M1, M2, M3) with mlx-whisper and voxtral-mlx
+- **Medium and large-v3 models** (we only tested base and small so far)
+- **Longer audio files** or domain-specific audio (medical, legal, call center)
+- **Other languages** beyond English and French

Dockerfile

@@ -1,83 +1,75 @@
-FROM nvidia/cuda:12.9.1-cudnn-devel-ubuntu24.04
+FROM ghcr.io/astral-sh/uv:0.10.4 AS uvbin
 
+# --- MARK: Builder Stage
+FROM nvidia/cuda:12.9.1-cudnn-devel-ubuntu24.04 AS builder-gpu
 ENV DEBIAN_FRONTEND=noninteractive
 ENV PYTHONUNBUFFERED=1
 
 WORKDIR /app
 
-ARG EXTRAS
-ARG HF_PRECACHE_DIR
-ARG HF_TKN_FILE
+RUN apt-get update && \
+  apt-get install -y --no-install-recommends \
+  build-essential \
+  python3-dev && \
+  rm -rf /var/lib/apt/lists/*
+
+# Install UV and set up the environment 
+COPY --from=uvbin /uv /uvx /bin/
+
+ENV UV_COMPILE_BYTECODE=1 UV_LINK_MODE=copy UV_NO_DEV=1
+ENV UV_PYTHON_PREFERENCE=only-managed
+ENV UV_PYTHON_INSTALL_DIR=/python
+
+RUN uv python install 3.12
+
+# Install dependencies first to leverage caching
+ARG EXTRAS=cu129
+COPY pyproject.toml uv.lock /app/
+RUN set -eux; \
+  set --; \
+  for extra in $(echo "${EXTRAS:-}" | tr ',' ' '); do \
+  set -- "$@" --extra "$extra"; \
+  done; \
+  uv sync --frozen --no-install-project --no-editable --no-cache "$@"
+
+# Copy the source code and install the package only
+COPY whisperlivekit /app/whisperlivekit
+RUN set -eux; \
+  set --; \
+  for extra in $(echo "${EXTRAS:-}" | tr ',' ' '); do \
+  set -- "$@" --extra "$extra"; \
+  done; \
+  uv sync --frozen --no-editable --no-cache "$@"
+
+# --- MARK: Runtime Stage 
+FROM nvidia/cuda:12.9.1-cudnn-runtime-ubuntu24.04
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+WORKDIR /app
 
 RUN apt-get update && \
-    apt-get install -y --no-install-recommends \
-        python3 \
-        python3-pip \
-        python3-venv \
-        ffmpeg \
-        git \
-        build-essential \
-        python3-dev \
-        ca-certificates && \
-    rm -rf /var/lib/apt/lists/*
+  apt-get install -y --no-install-recommends \
+  ffmpeg &&\
+  rm -rf /var/lib/apt/lists/*
 
-RUN python3 -m venv /opt/venv
-ENV PATH="/opt/venv/bin:$PATH"
+# Copy UV binaries
+COPY --from=uvbin /uv /uvx /bin/
 
-# timeout/retries for large torch wheels
-RUN pip3 install --upgrade pip setuptools wheel && \
-    pip3 --disable-pip-version-check install --timeout=120 --retries=5 \
-        --index-url https://download.pytorch.org/whl/cu129 \
-        torch torchaudio \
-    || (echo "Initial install failed — retrying with extended timeout..." && \
-        pip3 --disable-pip-version-check install --timeout=300 --retries=3 \
-            --index-url https://download.pytorch.org/whl/cu129 \
-            torch torchvision torchaudio)
+# Copy the Python version
+COPY --from=builder-gpu --chown=python:python /python /python
 
-COPY . .
-
-# Install WhisperLiveKit directly, allowing for optional dependencies
-RUN if [ -n "$EXTRAS" ]; then \
-      echo "Installing with extras: [$EXTRAS]"; \
-      pip install --no-cache-dir whisperlivekit[$EXTRAS]; \
-    else \
-      echo "Installing base package only"; \
-      pip install --no-cache-dir whisperlivekit; \
-    fi
-
-# In-container caching for Hugging Face models by: 
-# A) Make the cache directory persistent via an anonymous volume.
-#    Note: This only persists for a single, named container. This is 
-#          only for convenience at de/test stage. 
-#          For prod, it is better to use a named volume via host mount/k8s.
-VOLUME ["/root/.cache/huggingface/hub"]
-
-
-# or
-# B) Conditionally copy a local pre-cache from the build context to the 
-#    container's cache via the HF_PRECACHE_DIR build-arg.
-#    WARNING: This will copy ALL files in the pre-cache location.
-
-# Conditionally copy a cache directory if provided
-RUN if [ -n "$HF_PRECACHE_DIR" ]; then \
-      echo "Copying Hugging Face cache from $HF_PRECACHE_DIR"; \
-      mkdir -p /root/.cache/huggingface/hub && \
-      cp -r $HF_PRECACHE_DIR/* /root/.cache/huggingface/hub; \
-    else \
-      echo "No local Hugging Face cache specified, skipping copy"; \
-    fi
-
-# Conditionally copy a Hugging Face token if provided. Useful for Diart backend (pyannote audio models)
-RUN if [ -n "$HF_TKN_FILE" ]; then \
-      echo "Copying Hugging Face token from $HF_TKN_FILE"; \
-      mkdir -p /root/.cache/huggingface && \
-      cp $HF_TKN_FILE /root/.cache/huggingface/token; \
-    else \
-      echo "No Hugging Face token file specified, skipping token setup"; \
-    fi
+# Copy the virtual environment with all dependencies installed
+COPY --from=builder-gpu /app/.venv /app/.venv
 
 EXPOSE 8000
 
+ENV PATH="/app/.venv/bin:$PATH"
+ENV UV_PYTHON_DOWNLOADS=0
+
+HEALTHCHECK --interval=30s --timeout=5s --start-period=120s --retries=3 \
+  CMD python -c "import urllib.request; urllib.request.urlopen('http://localhost:8000/')" || exit 1
+
 ENTRYPOINT ["whisperlivekit-server", "--host", "0.0.0.0"]
 
 CMD ["--model", "medium"]

Dockerfile.cpu

@@ -1,61 +1,76 @@
-FROM python:3.13-slim
+FROM ghcr.io/astral-sh/uv:0.10.4 AS uvbin
 
+# --- MARK: Builder Stage
+FROM debian:bookworm-slim AS builder-cpu
 ENV DEBIAN_FRONTEND=noninteractive
 ENV PYTHONUNBUFFERED=1
 
 WORKDIR /app
 
-ARG EXTRAS
-ARG HF_PRECACHE_DIR
-ARG HF_TKN_FILE
+RUN apt-get update && \
+  apt-get install -y --no-install-recommends \
+  build-essential \
+  python3-dev && \
+  rm -rf /var/lib/apt/lists/*
+
+# Install UV and set up the environment 
+COPY --from=uvbin /uv /uvx /bin/
+
+ENV UV_COMPILE_BYTECODE=1 UV_LINK_MODE=copy UV_NO_DEV=1
+ENV UV_PYTHON_PREFERENCE=only-managed
+ENV UV_PYTHON_INSTALL_DIR=/python
+
+RUN uv python install 3.12
+
+# Install dependencies first to leverage caching
+ARG EXTRAS=cpu
+COPY pyproject.toml uv.lock /app/
+RUN set -eux; \
+  set --; \
+  for extra in $(echo "${EXTRAS:-}" | tr ',' ' '); do \
+  set -- "$@" --extra "$extra"; \
+  done; \
+  uv sync --frozen --no-install-project --no-editable --no-cache "$@"
+
+# Copy the source code and install the package only
+COPY whisperlivekit /app/whisperlivekit
+RUN set -eux; \
+  set --; \
+  for extra in $(echo "${EXTRAS:-}" | tr ',' ' '); do \
+  set -- "$@" --extra "$extra"; \
+  done; \
+  uv sync --frozen --no-editable --no-cache "$@"
+
+# --- MARK: Runtime Stage 
+FROM debian:bookworm-slim
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+WORKDIR /app
 
 RUN apt-get update && \
-    apt-get install -y --no-install-recommends \
-        ffmpeg \
-        git \
-        build-essential \
-        python3-dev && \
-    rm -rf /var/lib/apt/lists/*
+  apt-get install -y --no-install-recommends \
+  ffmpeg &&\
+  rm -rf /var/lib/apt/lists/*
 
-# Install CPU-only PyTorch
-RUN pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
+# Copy UV binaries
+COPY --from=uvbin /uv /uvx /bin/
 
-COPY . .
+# Copy the Python version
+COPY --from=builder-cpu --chown=python:python /python /python
 
-# Install WhisperLiveKit directly, allowing for optional dependencies
-RUN if [ -n "$EXTRAS" ]; then \
-      echo "Installing with extras: [$EXTRAS]"; \
-      pip install --no-cache-dir whisperlivekit[$EXTRAS]; \
-    else \
-      echo "Installing base package only"; \
-      pip install --no-cache-dir whisperlivekit; \
-    fi
+# Copy the virtual environment with all dependencies installed
+COPY --from=builder-cpu /app/.venv /app/.venv
 
-# Enable in-container caching for Hugging Face models
-VOLUME ["/root/.cache/huggingface/hub"]
-
-# Conditionally copy a local pre-cache from the build context
-RUN if [ -n "$HF_PRECACHE_DIR" ]; then \
-      echo "Copying Hugging Face cache from $HF_PRECACHE_DIR"; \
-      mkdir -p /root/.cache/huggingface/hub && \
-      cp -r $HF_PRECACHE_DIR/* /root/.cache/huggingface/hub; \
-    else \
-      echo "No local Hugging Face cache specified, skipping copy"; \
-    fi
-
-# Conditionally copy a Hugging Face token if provided
-RUN if [ -n "$HF_TKN_FILE" ]; then \
-      echo "Copying Hugging Face token from $HF_TKN_FILE"; \
-      mkdir -p /root/.cache/huggingface && \
-      cp $HF_TKN_FILE /root/.cache/huggingface/token; \
-    else \
-      echo "No Hugging Face token file specified, skipping token setup"; \
-    fi
-    
-# Expose port for the transcription server
 EXPOSE 8000
 
+ENV PATH="/app/.venv/bin:$PATH"
+ENV UV_PYTHON_DOWNLOADS=0
+
+HEALTHCHECK --interval=30s --timeout=5s --start-period=120s --retries=3 \
+  CMD python -c "import urllib.request; urllib.request.urlopen('http://localhost:8000/')" || exit 1
+
 ENTRYPOINT ["whisperlivekit-server", "--host", "0.0.0.0"]
 
 # Default args - you might want to use a smaller model for CPU
-CMD ["--model", "tiny"]
+CMD ["--model", "tiny"]

README.md

@@ -1,28 +1,32 @@
-<h1 align="center">WhisperLiveKit</h1>
+<h1 align="center">WLK</h1>
+<p align="center"><b>WhisperLiveKit: Ultra-low-latency, self-hosted speech-to-text with speaker identification</b></p>
+
 
 <p align="center">
 <img src="https://raw.githubusercontent.com/QuentinFuxa/WhisperLiveKit/refs/heads/main/demo.png" alt="WhisperLiveKit Demo" width="730">
 </p>
 
-<p align="center"><b>Real-time, Fully Local Speech-to-Text with Speaker Identification</b></p>
 
 <p align="center">
 <a href="https://pypi.org/project/whisperlivekit/"><img alt="PyPI Version" src="https://img.shields.io/pypi/v/whisperlivekit?color=g"></a>
 <a href="https://pepy.tech/project/whisperlivekit"><img alt="PyPI Downloads" src="https://static.pepy.tech/personalized-badge/whisperlivekit?period=total&units=international_system&left_color=grey&right_color=brightgreen&left_text=installations"></a>
 <a href="https://pypi.org/project/whisperlivekit/"><img alt="Python Versions" src="https://img.shields.io/badge/python-3.9--3.15-dark_green"></a>
+<a href="https://huggingface.co/qfuxa/whisper-base-french-lora">
+  <img alt="Hugging Face Weights" src="https://img.shields.io/badge/🤗-Hugging%20Face%20Weights-yellow" />
+</a>
 <a href="https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/LICENSE"><img alt="License" src="https://img.shields.io/badge/License-Apache 2.0-dark_green"></a>
 </p>
 
 
-Real-time transcription directly to your browser, with a ready-to-use backend+server and a simple frontend.
+### Powered by Leading Research:
 
-#### Powered by Leading Research:
-
-- Simul-[Whisper](https://github.com/backspacetg/simul_whisper)/[Streaming](https://github.com/ufal/SimulStreaming) (SOTA 2025) - Ultra-low latency transcription using [AlignAtt policy](https://arxiv.org/pdf/2305.11408)
+**See the interactive playground in [this repo](https://github.com/QuentinFuxa/streamlit-d3-network) to explore how AlignAtt works**
+- Simul-[Whisper](https://arxiv.org/pdf/2406.10052)/[Streaming](https://arxiv.org/abs/2506.17077) (SOTA 2025) - Ultra-low latency transcription using [AlignAtt policy](https://arxiv.org/pdf/2305.11408). 
 - [NLLW](https://github.com/QuentinFuxa/NoLanguageLeftWaiting) (2025), based on [distilled](https://huggingface.co/entai2965/nllb-200-distilled-600M-ctranslate2) [NLLB](https://arxiv.org/abs/2207.04672) (2022, 2024) - Simulatenous translation from & to 200 languages.
 - [WhisperStreaming](https://github.com/ufal/whisper_streaming) (SOTA 2023) - Low latency transcription using [LocalAgreement policy](https://www.isca-archive.org/interspeech_2020/liu20s_interspeech.pdf)
 - [Streaming Sortformer](https://arxiv.org/abs/2507.18446) (SOTA 2025) - Advanced real-time speaker diarization
 - [Diart](https://github.com/juanmc2005/diart) (SOTA 2021) - Real-time speaker diarization
+- [Voxtral Mini](https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602) (2025) - 4B-parameter multilingual speech model by Mistral AI
 - [Silero VAD](https://github.com/snakers4/silero-vad) (2024) - Enterprise-grade Voice Activity Detection
 
 
@@ -51,9 +55,11 @@ pip install whisperlivekit
 2. **Open your browser** and navigate to `http://localhost:8000`. Start speaking and watch your words appear in real-time!
 
 
-> - See [tokenizer.py](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/simul_whisper/whisper/tokenizer.py) for the list of all available languages.
-> - For HTTPS requirements, see the **Parameters** section for SSL configuration options.
+> - See [here](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/simul_whisper/whisper/tokenizer.py) for the list of all available languages.
+> - Check the [troubleshooting guide](docs/troubleshooting.md) for step-by-step fixes collected from recent GPU setup/env issues.
 > - The CLI entry point is exposed as both `wlk` and `whisperlivekit-server`; they are equivalent.
+> - For HTTPS requirements, see the **Parameters** section for SSL configuration options.
+
 
 #### Use it to capture audio from web pages.
 
@@ -67,19 +73,61 @@ Go to `chrome-extension` for instructions.
 
 #### Optional Dependencies
 
-| Optional | `pip install` |
-|-----------|-------------|
-| **Windows/Linux optimizations** | `faster-whisper` |
-| **Apple Silicon optimizations** | `mlx-whisper` |
-| **Translation** | `nllw` |
-| **Speaker diarization** | `git+https://github.com/NVIDIA/NeMo.git@main#egg=nemo_toolkit[asr]` |
-| OpenAI API | `openai` |
-| *[Not recommanded]*  Speaker diarization with Diart | `diart` |
+| Feature | `uv sync` | `pip install -e` |
+|-----------|-------------|-------------|
+| **Apple Silicon MLX Whisper backend** | `uv sync --extra mlx-whisper` | `pip install -e ".[mlx-whisper]"` |
+| **Voxtral (MLX backend, Apple Silicon)** | `uv sync --extra voxtral-mlx` | `pip install -e ".[voxtral-mlx]"` |
+| **CPU PyTorch stack** | `uv sync --extra cpu` | `pip install -e ".[cpu]"` |
+| **CUDA 12.9 PyTorch stack** | `uv sync --extra cu129` | `pip install -e ".[cu129]"` |
+| **Translation** | `uv sync --extra translation` | `pip install -e ".[translation]"` |
+| **Sentence tokenizer** | `uv sync --extra sentence_tokenizer` | `pip install -e ".[sentence_tokenizer]"` |
+| **Voxtral (HF backend)** | `uv sync --extra voxtral-hf` | `pip install -e ".[voxtral-hf]"` |
+| **Speaker diarization (Sortformer / NeMo)** | `uv sync --extra diarization-sortformer` | `pip install -e ".[diarization-sortformer]"` |
+| *[Not recommended]* Speaker diarization with Diart | `uv sync --extra diarization-diart` | `pip install -e ".[diarization-diart]"` |
 
-See  **Parameters & Configuration** below on how to use them.
+Supported GPU profiles:
+
+```bash
+# Profile A: Sortformer diarization
+uv sync --extra cu129 --extra diarization-sortformer
+
+# Profile B: Voxtral HF + translation
+uv sync --extra cu129 --extra voxtral-hf --extra translation
+```
+
+`voxtral-hf` and `diarization-sortformer` are intentionally incompatible extras and must be installed in separate environments.
+
+See **Parameters & Configuration** below on how to use them.
+
+<p align="center">
+<img src="benchmark_scatter.png" alt="Speed vs Accuracy tradeoff" width="700">
+</p>
+
+See **[BENCHMARK.md](BENCHMARK.md)** for the full benchmark with tables, model size comparison, and more.
+We are actively looking for benchmark results on other hardware (NVIDIA GPUs, different Apple Silicon chips, cloud instances). If you run the benchmarks on your machine, please share your results via an issue or PR!
 
 
 
+### Voxtral Backend
+
+WhisperLiveKit supports [Voxtral Mini](https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602),
+a 4B-parameter speech model from Mistral AI that natively handles 100+ languages with automatic
+language detection. Whisper also supports auto-detection (`--language auto`), but Voxtral's per-chunk
+detection is more reliable and does not bias towards English.
+
+```bash
+# Apple Silicon (native MLX, recommended)
+pip install -e ".[voxtral-mlx]"
+wlk --backend voxtral-mlx
+
+# Linux/GPU (HuggingFace transformers)
+pip install transformers torch
+wlk --backend voxtral
+```
+
+Voxtral uses its own streaming policy and does not use LocalAgreement or SimulStreaming.
+See [BENCHMARK.md](BENCHMARK.md) for performance numbers.
+
 ### Usage Examples
 
 **Command-line Interface**: Start the transcription server with various options:
@@ -88,19 +136,24 @@ See  **Parameters & Configuration** below on how to use them.
 # Large model and translate from french to danish
 wlk --model large-v3 --language fr --target-language da
 
-# Diarization and server listening on */80 
+# Diarization and server listening on */80
 wlk --host 0.0.0.0 --port 80 --model medium --diarization --language fr
+
+# Voxtral multilingual (auto-detects language)
+wlk --backend voxtral-mlx
 ```
 
 
 **Python API Integration**: Check [basic_server](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/basic_server.py) for a more complete example of how to use the functions and classes.
 
 ```python
-from whisperlivekit import TranscriptionEngine, AudioProcessor, parse_args
+import asyncio
+from contextlib import asynccontextmanager
+
 from fastapi import FastAPI, WebSocket, WebSocketDisconnect
 from fastapi.responses import HTMLResponse
-from contextlib import asynccontextmanager
-import asyncio
+
+from whisperlivekit import AudioProcessor, TranscriptionEngine, parse_args
 
 transcription_engine = None
 
@@ -139,15 +192,15 @@ async def websocket_endpoint(websocket: WebSocket):
 
 | Parameter | Description | Default |
 |-----------|-------------|---------|
-| `--model` | Whisper model size. List and recommandations [here](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/docs/available_models.md) | `small` |
-| `--model-path` | Local .pt file/directory **or** Hugging Face repo ID containing the Whisper model. Overrides `--model`. Recommandations [here](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/docs/models_compatible_formats.md) | `None` |
-| `--language` | List [here](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/whisper/tokenizer.py). If you use `auto`, the model attempts to detect the language automatically, but it tends to bias towards English. | `auto` |
-| `--target-language` | If sets, translates using [NLLW](https://github.com/QuentinFuxa/NoLanguageLeftWaiting). [200 languages available](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/docs/supported_languages.md). If you want to translate to english, you can also use `--direct-english-translation`. The STT model will try to directly output the translation. | `None` |
+| `--model` | Whisper model size. List and recommandations [here](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/docs/default_and_custom_models.md) | `small` |
+| `--model-path` | Local .pt file/directory **or** Hugging Face repo ID containing the Whisper model. Overrides `--model`. Recommandations [here](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/docs/default_and_custom_models.md) | `None` |
+| `--language` | List [here](docs/supported_languages.md). If you use `auto`, the model attempts to detect the language automatically, but it tends to bias towards English. | `auto` |
+| `--target-language` | If sets, translates using [NLLW](https://github.com/QuentinFuxa/NoLanguageLeftWaiting). [200 languages available](docs/supported_languages.md). If you want to translate to english, you can also use `--direct-english-translation`. The STT model will try to directly output the translation. | `None` |
 | `--diarization` | Enable speaker identification | `False` |
 | `--backend-policy` | Streaming strategy: `1`/`simulstreaming` uses AlignAtt SimulStreaming, `2`/`localagreement` uses the LocalAgreement policy | `simulstreaming` |
-| `--backend` | Whisper implementation selector. `auto` picks MLX on macOS (if installed), otherwise Faster-Whisper, otherwise vanilla Whisper. You can also force `mlx-whisper`, `faster-whisper`, `whisper`, or `openai-api` (LocalAgreement only) | `auto` |
-| `--no-vac` | Disable Voice Activity Controller | `False` |
-| `--no-vad` | Disable Voice Activity Detection | `False` |
+| `--backend` | ASR backend selector. `auto` picks MLX on macOS (if installed), otherwise Faster-Whisper, otherwise vanilla Whisper. Options: `mlx-whisper`, `faster-whisper`, `whisper`, `openai-api` (LocalAgreement only), `voxtral-mlx` (Apple Silicon), `voxtral` (HuggingFace) | `auto` |
+| `--no-vac` | Disable Voice Activity Controller. NOT ADVISED | `False` |
+| `--no-vad` | Disable Voice Activity Detection. NOT ADVISED | `False` |
 | `--warmup-file` | Audio file path for model warmup | `jfk.wav` |
 | `--host` | Server host address | `localhost` |
 | `--port` | Server port | `8000` |
@@ -155,6 +208,7 @@ async def websocket_endpoint(websocket: WebSocket):
 | `--ssl-keyfile` | Path to the SSL private key file (for HTTPS support) | `None` |
 | `--forwarded-allow-ips` | Ip or Ips allowed to reverse proxy the whisperlivekit-server. Supported types are  IP Addresses (e.g. 127.0.0.1), IP Networks (e.g. 10.100.0.0/16), or Literals (e.g. /path/to/socket.sock) | `None` |
 | `--pcm-input` | raw PCM (s16le) data is expected as input and FFmpeg will be bypassed. Frontend will use AudioWorklet instead of MediaRecorder | `False` |
+| `--lora-path` | Path or Hugging Face repo ID for LoRA adapter weights (e.g., `qfuxa/whisper-base-french-lora`). Only works with native Whisper backend (`--backend whisper`) | `None` |
 
 | Translation options | Description | Default |
 |-----------|-------------|---------|
@@ -164,7 +218,7 @@ async def websocket_endpoint(websocket: WebSocket):
 | Diarization options | Description | Default |
 |-----------|-------------|---------|
 | `--diarization-backend` |  `diart` or `sortformer` | `sortformer` |
-| `--disable-punctuation-split` |  Disable punctuation based splits. See #214 | `False` |
+| `--disable-punctuation-split` | [NOT FUNCTIONAL IN 0.2.15 / 0.2.16] Disable punctuation based splits. See #214 | `False` |
 | `--segmentation-model` | Hugging Face model ID for Diart segmentation model. [Available models](https://github.com/juanmc2005/diart/tree/main?tab=readme-ov-file#pre-trained-models) | `pyannote/segmentation-3.0` |
 | `--embedding-model` | Hugging Face model ID for Diart embedding model. [Available models](https://github.com/juanmc2005/diart/tree/main?tab=readme-ov-file#pre-trained-models) | `speechbrain/spkrec-ecapa-voxceleb` |
 
@@ -182,8 +236,7 @@ async def websocket_endpoint(websocket: WebSocket):
 | `--never-fire` | Never truncate incomplete words | `False` |
 | `--init-prompt` | Initial prompt for the model | `None` |
 | `--static-init-prompt` | Static prompt that doesn't scroll | `None` |
-| `--max-context-tokens` | Maximum context tokens | `None` |
-| `--preload-model-count` | Optional. Number of models to preload in memory to speed up loading (set up to the expected number of concurrent users) | `1` |
+| `--max-context-tokens` | Maximum context tokens | Depends on model used, but usually 448. |
 
 
 
@@ -242,7 +295,7 @@ docker run --gpus all -p 8000:8000 --name wlk wlk
 
 **CPU only:**
 ```bash
-docker build -f Dockerfile.cpu -t wlk .
+docker build -f Dockerfile.cpu -t wlk --build-arg EXTRAS="cpu" .
 docker run -p 8000:8000 --name wlk wlk
 ```
 
@@ -254,6 +307,18 @@ docker run -p 8000:8000 --name wlk wlk
 docker run --gpus all -p 8000:8000 --name wlk wlk --model large-v3 --language fr
 ```
 
+**Compose (recommended for cache + token wiring):**
+```bash
+# GPU Sortformer profile
+docker compose up --build wlk-gpu-sortformer
+
+# GPU Voxtral profile
+docker compose up --build wlk-gpu-voxtral
+
+# CPU service
+docker compose up --build wlk-cpu
+```
+
 ### Memory Requirements
 - **Large models**: Ensure your Docker runtime has sufficient memory allocated
 
@@ -261,9 +326,34 @@ docker run --gpus all -p 8000:8000 --name wlk wlk --model large-v3 --language fr
 #### Customization
 
 - `--build-arg` Options:
-  - `EXTRAS="whisper-timestamped"` - Add extras to the image's installation (no spaces). Remember to set necessary container options!
-  - `HF_PRECACHE_DIR="./.cache/"` - Pre-load a model cache for faster first-time start
-  - `HF_TKN_FILE="./token"` - Add your Hugging Face Hub access token to download gated models
+  - `EXTRAS="cu129,diarization-sortformer"` - GPU Sortformer profile extras.
+  - `EXTRAS="cu129,voxtral-hf,translation"` - GPU Voxtral profile extras.
+  - `EXTRAS="cpu,diarization-diart,translation"` - CPU profile extras.
+  - Hugging Face cache + token are configured in `compose.yml` using a named volume and `HF_TKN_FILE` (default: `./token`).
 
-## 🔮 Use Cases
+## Testing & Benchmarks
+
+WhisperLiveKit includes a unit test suite and an offline benchmark harness.
+
+```bash
+# Install test dependencies
+pip install -e ".[test]"
+
+# Run unit tests (no model download required)
+pytest tests/ -v
+
+# Benchmark a single backend
+python test_backend_offline.py --backend faster-whisper --no-realtime
+
+# Benchmark all installed backends
+python test_backend_offline.py --benchmark --no-realtime
+
+# Export benchmark results as JSON
+python test_backend_offline.py --benchmark --no-realtime --json results.json
+```
+
+See [BENCHMARK.md](BENCHMARK.md) for a full comparison of backends, policies, WER, speed, and
+timestamp accuracy on Apple Silicon.
+
+## Use Cases
 Capture discussions in real-time for meeting transcription, help hearing-impaired users follow conversations through accessibility tools, transcribe podcasts or videos automatically for content creation, transcribe support calls with speaker identification for customer service...

ReadmeJP.md

@@ -1,258 +0,0 @@
-<h1 align="center">WhisperLiveKit</h1>
-
-<p align="center">
-<img src="https://raw.githubusercontent.com/QuentinFuxa/WhisperLiveKit/refs/heads/main/demo.png" alt="WhisperLiveKit Demo" width="730">
-</p>
-
-<p align="center"><b>話者識別機能付き、リアルタイム、完全ローカルな音声テキスト変換</b></p>
-
-<p align="center">
-<a href="https://pypi.org/project/whisperlivekit/"><img alt="PyPI Version" src="https://img.shields.io/pypi/v/whisperlivekit?color=g"></a>
-<a href="https://pepy.tech/project/whisperlivekit"><img alt="PyPI Downloads" src="https://static.pepy.tech/personalized-badge/whisperlivekit?period=total&units=international_system&left_color=grey&right_color=brightgreen&left_text=installations"></a>
-<a href="https://pypi.org/project/whisperlivekit/"><img alt="Python Versions" src="https://img.shields.io/badge/python-3.9--3.13-dark_green"></a>
-<a href="https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/LICENSE"><img alt="License" src="https://img.shields.io/badge/License-MIT/Dual Licensed-dark_green"></a>
-</p>
-
-すぐに使えるバックエンド+サーバーとシンプルなフロントエンドで、リアルタイムの音声文字起こしをブラウザに直接提供します。✨
-
-#### 主要な研究による技術：
-
-- [SimulStreaming](https://github.com/ufal/SimulStreaming) (SOTA 2025) - AlignAttポリシーによる超低遅延文字起こし
-- [WhisperStreaming](https://github.com/ufal/whisper_streaming) (SOTA 2023) - LocalAgreementポリシーによる低遅延文字起こし
-- [Streaming Sortformer](https://arxiv.org/abs/2507.18446) (SOTA 2025) - 高度なリアルタイム話者ダイアライゼーション
-- [Diart](https://github.com/juanmc2005/diart) (SOTA 2021) - リアルタイム話者ダイアライゼーション
-- [Silero VAD](https://github.com/snakers4/silero-vad) (2024) - エンタープライズグレードの音声区間検出
-
-> **なぜ各音声バッチで単純なWhisperモデルを実行しないのか？** Whisperは完全な発話向けに設計されており、リアルタイムのチャンク向けではありません。小さなセグメントを処理するとコンテキストが失われ、単語が音節の途中で途切れ、質の悪い文字起こしになります。WhisperLiveKitは、インテリジェントなバッファリングとインクリメンタルな処理のために、最先端の同時音声研究を利用しています。
-
-### アーキテクチャ
-
-<img alt="Architecture" src="https://raw.githubusercontent.com/QuentinFuxa/WhisperLiveKit/refs/heads/main/architecture.png" />
-
-*バックエンドは複数の同時ユーザーをサポートします。音声が検出されない場合、音声区間検出がオーバーヘッドを削減します。*
-
-### インストールとクイックスタート
-
-```bash
-pip install whisperlivekit
-```
-
->  **FFmpegが必要です** WhisperLiveKitを使用する前にインストールする必要があります。
->
-> | OS | インストール方法 |
-> |-----------|-------------|
->  | Ubuntu/Debian | `sudo apt install ffmpeg` |
-> | MacOS | `brew install ffmpeg` |
-> | Windows | https://ffmpeg.org/download.html から.exeをダウンロードし、PATHに追加 |
-
-#### クイックスタート
-1. **文字起こしサーバーを起動します:**
-   ```bash
-   whisperlivekit-server --model base --language en
-   ```
-
-2. **ブラウザを開き** `http://localhost:8000` にアクセスします。話し始めると、あなたの言葉がリアルタイムで表示されます！
-
-
-> - 利用可能なすべての言語のリストについては、[tokenizer.py](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/simul_whisper/whisper/tokenizer.py) を参照してください。
-> - HTTPSの要件については、**パラメータ**セクションのSSL設定オプションを参照してください。
-
-#### オプションの依存関係
-
-| オプション | `pip install` |
-|-----------|-------------|
-| **Sortformerによる話者ダイアライゼーション** | `git+https://github.com/NVIDIA/NeMo.git@main#egg=nemo_toolkit[asr]` |
-| Diartによる話者ダイアライゼーション | `diart` |
-| オリジナルのWhisperバックエンド | `whisper` |
-| タイムスタンプ改善バックエンド | `whisper-timestamped` |
-| Apple Silicon最適化バックエンド | `mlx-whisper` |
-| OpenAI APIバックエンド | `openai` |
-
-それらの使用方法については、以下の**パラメータと設定**を参照してください。
-
-### 使用例
-
-**コマンドラインインターフェース**: 様々なオプションで文字起こしサーバーを起動します:
-
-```bash
-# デフォルト(small)より良いモデルを使用
-whisperlivekit-server --model large-v3
-
-# ダイアライゼーションと言語を指定した高度な設定
-whisperlivekit-server --host 0.0.0.0 --port 8000 --model medium --diarization --language fr
-```
-
-**Python API連携**: 関数やクラスの使用方法のより完全な例については、[basic_server](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/basic_server.py) を確認してください。
-
-```python
-from whisperlivekit import TranscriptionEngine, AudioProcessor, parse_args
-from fastapi import FastAPI, WebSocket, WebSocketDisconnect
-from fastapi.responses import HTMLResponse
-from contextlib import asynccontextmanager
-import asyncio
-
-transcription_engine = None
-
-@asynccontextmanager
-async def lifespan(app: FastAPI):
-    global transcription_engine
-    transcription_engine = TranscriptionEngine(model="medium", diarization=True, lan="en")
-    yield
-
-app = FastAPI(lifespan=lifespan)
-
-async def handle_websocket_results(websocket: WebSocket, results_generator):
-    async for response in results_generator:
-        await websocket.send_json(response)
-    await websocket.send_json({"type": "ready_to_stop"})
-
-@app.websocket("/asr")
-async def websocket_endpoint(websocket: WebSocket):
-    global transcription_engine
-
-    # 接続ごとに新しいAudioProcessorを作成し、共有エンジンを渡す
-    audio_processor = AudioProcessor(transcription_engine=transcription_engine)
-    results_generator = await audio_processor.create_tasks()
-    results_task = asyncio.create_task(handle_websocket_results(websocket, results_generator))
-    await websocket.accept()
-    while True:
-        message = await websocket.receive_bytes()
-        await audio_processor.process_audio(message)
-```
-
-**フロントエンド実装**: パッケージにはHTML/JavaScript実装が[ここ](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/web/live_transcription.html)に含まれています。`from whisperlivekit import get_web_interface_html` & `page = get_web_interface_html()` を使ってインポートすることもできます。
-
-
-## パラメータと設定
-
-重要なパラメータのリストを変更できます。しかし、何を*変更すべき*でしょうか？
-- `--model` サイズ。リストと推奨事項は[こちら](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/available_models.md)
-- `--language`。リストは[こちら](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/simul_whisper/whisper/tokenizer.py)。`auto`を使用すると、モデルは自動的に言語を検出しようとしますが、英語に偏る傾向があります。
-- `--backend`？ `simulstreaming`が正しく動作しない場合や、デュアルライセンス要件を避けたい場合は`--backend faster-whisper`に切り替えることができます。
-- `--warmup-file`、もしあれば
-- `--host`, `--port`, `--ssl-certfile`, `--ssl-keyfile`、サーバーをセットアップする場合
-- `--diarization`、使用したい場合。
-
-残りは推奨しません。しかし、以下があなたのオプションです。
-
-| パラメータ | 説明 | デフォルト |
-|-----------|-------------|---------|
-| `--model` | Whisperモデルのサイズ。 | `small` |
-| `--language` | ソース言語コードまたは`auto` | `auto` |
-| `--task` | `transcribe`または`translate` | `transcribe` |
-| `--backend` | 処理バックエンド | `simulstreaming` |
-| `--min-chunk-size` | 最小音声チャンクサイズ（秒） | `1.0` |
-| `--no-vac` | 音声アクティビティコントローラーを無効化 | `False` |
-| `--no-vad` | 音声区間検出を無効化 | `False` |
-| `--warmup-file` | モデルのウォームアップ用音声ファイルパス | `jfk.wav` |
-| `--host` | サーバーホストアドレス | `localhost` |
-| `--port` | サーバーポート | `8000` |
-| `--ssl-certfile` | SSL証明書ファイルへのパス（HTTPSサポート用） | `None` |
-| `--ssl-keyfile` | SSL秘密鍵ファイルへのパス（HTTPSサポート用） | `None` |
-
-
-| WhisperStreamingバックエンドオプション | 説明 | デフォルト |
-|-----------|-------------|---------|
-| `--confidence-validation` | 高速な検証のために信頼スコアを使用 | `False` |
-| `--buffer_trimming` | バッファトリミング戦略（`sentence`または`segment`） | `segment` |
-
-
-| SimulStreamingバックエンドオプション | 説明 | デフォルト |
-|-----------|-------------|---------|
-| `--frame-threshold` | AlignAttフレームしきい値（低いほど速く、高いほど正確） | `25` |
-| `--beams` | ビームサーチのビーム数（1 = 貪欲デコーディング） | `1` |
-| `--decoder` | デコーダタイプを強制（`beam`または`greedy`） | `auto` |
-| `--audio-max-len` | 最大音声バッファ長（秒） | `30.0` |
-| `--audio-min-len` | 処理する最小音声長（秒） | `0.0` |
-| `--cif-ckpt-path` | 単語境界検出用CIFモデルへのパス | `None` |
-| `--never-fire` | 未完了の単語を決して切り捨てない | `False` |
-| `--init-prompt` | モデルの初期プロンプト | `None` |
-| `--static-init-prompt` | スクロールしない静的プロンプト | `None` |
-| `--max-context-tokens` | 最大コンテキストトークン数 | `None` |
-| `--model-path` | .ptモデルファイルへの直接パス。見つからない場合はダウンロード | `./base.pt` |
-| `--preloaded-model-count` | オプション。メモリにプリロードするモデルの数（予想される同時ユーザー数まで設定） | `1` |
-
-| ダイアライゼーションオプション | 説明 | デフォルト |
-|-----------|-------------|---------|
-| `--diarization` | 話者識別を有効化 | `False` |
-| `--diarization-backend` | `diart`または`sortformer` | `sortformer` |
-| `--segmentation-model` | DiartセグメンテーションモデルのHugging FaceモデルID。[利用可能なモデル](https://github.com/juanmc2005/diart/tree/main?tab=readme-ov-file#pre-trained-models) | `pyannote/segmentation-3.0` |
-| `--embedding-model` | Diart埋め込みモデルのHugging FaceモデルID。[利用可能なモデル](https://github.com/juanmc2005/diart/tree/main?tab=readme-ov-file#pre-trained-models) | `speechbrain/spkrec-ecapa-voxceleb` |
-
-
-> Diartを使用したダイアライゼーションには、pyannote.audioモデルへのアクセスが必要です：
-> 1. `pyannote/segmentation`モデルの[ユーザー条件に同意](https://huggingface.co/pyannote/segmentation)
-> 2. `pyannote/segmentation-3.0`モデルの[ユーザー条件に同意](https://huggingface.co/pyannote/segmentation-3.0)
-> 3. `pyannote/embedding`モデルの[ユーザー条件に同意](https://huggingface.co/pyannote/embedding)
->4. HuggingFaceでログイン: `huggingface-cli login`
-
-### 🚀 デプロイガイド
-
-WhisperLiveKitを本番環境にデプロイするには：
-
-1. **サーバーセットアップ**: 本番用ASGIサーバーをインストールし、複数のワーカーで起動します
-   ```bash
-   pip install uvicorn gunicorn
-   gunicorn -k uvicorn.workers.UvicornWorker -w 4 your_app:app
-   ```
-
-2. **フロントエンド**: カスタマイズした`html`のバージョンをホストし、WebSocket接続が正しくポイントするようにします
-
-3. **Nginx設定** (本番環境で推奨):
-    ```nginx
-   server {
-       listen 80;
-       server_name your-domain.com;
-        location / {
-            proxy_pass http://localhost:8000;
-            proxy_set_header Upgrade $http_upgrade;
-            proxy_set_header Connection "upgrade";
-            proxy_set_header Host $host;
-    }}
-    ```
-
-4. **HTTPSサポート**: 安全なデプロイメントのために、WebSocket URLで "ws://" の代わりに "wss://" を使用します
-
-## 🐋 Docker
-
-GPUまたはCPUサポート付きでDockerを使用してアプリケーションを簡単にデプロイします。
-
-### 前提条件
-- Dockerがシステムにインストールされていること
-- GPUサポートの場合: NVIDIA Dockerランタイムがインストールされていること
-
-### クイックスタート
-
-**GPUアクセラレーション付き (推奨):**
-```bash
-docker build -t wlk .
-docker run --gpus all -p 8000:8000 --name wlk wlk
-```
-
-**CPUのみ:**
-```bash
-docker build -f Dockerfile.cpu -t wlk .
-docker run -p 8000:8000 --name wlk wlk
-```
-
-### 高度な使用法
-
-**カスタム設定:**
-```bash
-# カスタムモデルと言語の例
-docker run --gpus all -p 8000:8000 --name wlk wlk --model large-v3 --language fr
-```
-
-### メモリ要件
-- **大規模モデル**: Dockerランタイムに十分なメモリが割り当てられていることを確認してください
-
-
-#### カスタマイズ
-
-- `--build-arg` オプション:
-  - `EXTRAS="whisper-timestamped"` - イメージのインストールにエクストラを追加します（スペースなし）。必要なコンテナオプションを設定することを忘れないでください！
-  - `HF_PRECACHE_DIR="./.cache/"` - 初回起動を高速化するためにモデルキャッシュをプリロードします
-  - `HF_TKN_FILE="./token"` - ゲート付きモデルをダウンロードするためにHugging Face Hubアクセストークンを追加します
-
-## 🔮 ユースケース
-会議の文字起こしのためにリアルタイムで議論をキャプチャする、聴覚障害のあるユーザーがアクセシビリティツールを通じて会話を追うのを助ける、コンテンツ作成のためにポッドキャストやビデオを自動的に文字起こしする、カスタマーサービスのために話者識別付きでサポートコールを文字起こしする...

audio_tests/00_00_07_english_1_speaker.transcript.json

@@ -0,0 +1,97 @@
+[
+  {
+    "word": "This",
+    "start": 0.0,
+    "end": 0.24
+  },
+  {
+    "word": "is",
+    "start": 0.24,
+    "end": 0.56
+  },
+  {
+    "word": "a",
+    "start": 0.56,
+    "end": 0.76
+  },
+  {
+    "word": "transcription",
+    "start": 0.76,
+    "end": 1.32
+  },
+  {
+    "word": "test.",
+    "start": 1.32,
+    "end": 2.0
+  },
+  {
+    "word": "We",
+    "start": 2.4,
+    "end": 2.5
+  },
+  {
+    "word": "want",
+    "start": 2.5,
+    "end": 2.66
+  },
+  {
+    "word": "to",
+    "start": 2.66,
+    "end": 2.84
+  },
+  {
+    "word": "see",
+    "start": 2.84,
+    "end": 3.1
+  },
+  {
+    "word": "if",
+    "start": 3.1,
+    "end": 3.34
+  },
+  {
+    "word": "we",
+    "start": 3.34,
+    "end": 3.5
+  },
+  {
+    "word": "can",
+    "start": 3.5,
+    "end": 3.68
+  },
+  {
+    "word": "use",
+    "start": 3.68,
+    "end": 4.04
+  },
+  {
+    "word": "smaller",
+    "start": 4.04,
+    "end": 4.76
+  },
+  {
+    "word": "chunks.",
+    "start": 4.76,
+    "end": 5.16
+  },
+  {
+    "word": "What",
+    "start": 6.06,
+    "end": 6.32
+  },
+  {
+    "word": "do",
+    "start": 6.32,
+    "end": 6.44
+  },
+  {
+    "word": "you",
+    "start": 6.44,
+    "end": 6.58
+  },
+  {
+    "word": "think?",
+    "start": 6.58,
+    "end": 6.84
+  }
+]

audio_tests/00_00_16_french_1_speaker.transcript.json

@@ -0,0 +1,177 @@
+[
+  {
+    "word": "Ok,",
+    "start": 2.02,
+    "end": 2.38
+  },
+  {
+    "word": "là",
+    "start": 2.52,
+    "end": 2.58
+  },
+  {
+    "word": "c",
+    "start": 2.58,
+    "end": 2.74
+  },
+  {
+    "word": "'est",
+    "start": 2.74,
+    "end": 2.76
+  },
+  {
+    "word": "un",
+    "start": 2.76,
+    "end": 2.86
+  },
+  {
+    "word": "test,",
+    "start": 2.86,
+    "end": 3.2
+  },
+  {
+    "word": "on",
+    "start": 3.34,
+    "end": 3.34
+  },
+  {
+    "word": "veut",
+    "start": 3.34,
+    "end": 3.48
+  },
+  {
+    "word": "voir",
+    "start": 3.48,
+    "end": 3.86
+  },
+  {
+    "word": "si",
+    "start": 3.86,
+    "end": 4.14
+  },
+  {
+    "word": "ça",
+    "start": 4.14,
+    "end": 4.26
+  },
+  {
+    "word": "arrive",
+    "start": 4.26,
+    "end": 4.36
+  },
+  {
+    "word": "à",
+    "start": 4.36,
+    "end": 4.5
+  },
+  {
+    "word": "capté",
+    "start": 4.5,
+    "end": 4.78
+  },
+  {
+    "word": "le",
+    "start": 4.78,
+    "end": 4.9
+  },
+  {
+    "word": "silence.",
+    "start": 4.9,
+    "end": 5.44
+  },
+  {
+    "word": "Là",
+    "start": 9.24,
+    "end": 9.6
+  },
+  {
+    "word": "il",
+    "start": 9.6,
+    "end": 9.78
+  },
+  {
+    "word": "est",
+    "start": 9.78,
+    "end": 9.84
+  },
+  {
+    "word": "une",
+    "start": 9.84,
+    "end": 9.96
+  },
+  {
+    "word": "telle",
+    "start": 9.96,
+    "end": 10.12
+  },
+  {
+    "word": "seconde",
+    "start": 10.12,
+    "end": 10.38
+  },
+  {
+    "word": "de",
+    "start": 10.38,
+    "end": 10.48
+  },
+  {
+    "word": "silence",
+    "start": 10.48,
+    "end": 10.78
+  },
+  {
+    "word": "et",
+    "start": 10.78,
+    "end": 11.06
+  },
+  {
+    "word": "je",
+    "start": 11.06,
+    "end": 11.16
+  },
+  {
+    "word": "vous",
+    "start": 11.16,
+    "end": 11.32
+  },
+  {
+    "word": "parle.",
+    "start": 11.32,
+    "end": 11.68
+  },
+  {
+    "word": "Et",
+    "start": 13.28,
+    "end": 13.64
+  },
+  {
+    "word": "voilà,",
+    "start": 13.64,
+    "end": 13.96
+  },
+  {
+    "word": "allez",
+    "start": 14.36,
+    "end": 14.62
+  },
+  {
+    "word": "on",
+    "start": 14.62,
+    "end": 14.78
+  },
+  {
+    "word": "va",
+    "start": 14.78,
+    "end": 14.88
+  },
+  {
+    "word": "tester",
+    "start": 14.88,
+    "end": 15.06
+  },
+  {
+    "word": "ça.",
+    "start": 15.06,
+    "end": 15.36
+  }
+]

audio_tests/00_00_30_english_3_speakers.transcript.json

@@ -0,0 +1,382 @@
+[
+  {
+    "word": "Transcription",
+    "start": 0.0,
+    "end": 0.6
+  },
+  {
+    "word": "technology",
+    "start": 0.6,
+    "end": 1.24
+  },
+  {
+    "word": "has",
+    "start": 1.24,
+    "end": 1.5
+  },
+  {
+    "word": "improved",
+    "start": 1.5,
+    "end": 1.96
+  },
+  {
+    "word": "so",
+    "start": 1.96,
+    "end": 2.32
+  },
+  {
+    "word": "much",
+    "start": 2.32,
+    "end": 2.68
+  },
+  {
+    "word": "in",
+    "start": 2.68,
+    "end": 2.94
+  },
+  {
+    "word": "the",
+    "start": 2.94,
+    "end": 3.02
+  },
+  {
+    "word": "past",
+    "start": 3.02,
+    "end": 3.24
+  },
+  {
+    "word": "few",
+    "start": 3.24,
+    "end": 3.5
+  },
+  {
+    "word": "years.",
+    "start": 3.5,
+    "end": 3.96
+  },
+  {
+    "word": "Have",
+    "start": 4.56,
+    "end": 4.74
+  },
+  {
+    "word": "you",
+    "start": 4.74,
+    "end": 4.9
+  },
+  {
+    "word": "noticed",
+    "start": 4.9,
+    "end": 5.26
+  },
+  {
+    "word": "how",
+    "start": 5.26,
+    "end": 5.52
+  },
+  {
+    "word": "accurate",
+    "start": 5.52,
+    "end": 6.08
+  },
+  {
+    "word": "real",
+    "start": 6.08,
+    "end": 6.42
+  },
+  {
+    "word": "-time",
+    "start": 6.42,
+    "end": 6.74
+  },
+  {
+    "word": "speech",
+    "start": 6.74,
+    "end": 7.24
+  },
+  {
+    "word": "to",
+    "start": 7.24,
+    "end": 7.46
+  },
+  {
+    "word": "text",
+    "start": 7.46,
+    "end": 7.78
+  },
+  {
+    "word": "is",
+    "start": 7.78,
+    "end": 8.0
+  },
+  {
+    "word": "now?",
+    "start": 8.0,
+    "end": 8.3
+  },
+  {
+    "word": "Absolutely.",
+    "start": 8.7,
+    "end": 9.16
+  },
+  {
+    "word": "I",
+    "start": 10.04,
+    "end": 10.38
+  },
+  {
+    "word": "use",
+    "start": 10.38,
+    "end": 10.56
+  },
+  {
+    "word": "it",
+    "start": 10.56,
+    "end": 10.76
+  },
+  {
+    "word": "all",
+    "start": 10.76,
+    "end": 10.9
+  },
+  {
+    "word": "the",
+    "start": 10.9,
+    "end": 11.04
+  },
+  {
+    "word": "time",
+    "start": 11.04,
+    "end": 11.32
+  },
+  {
+    "word": "for",
+    "start": 11.32,
+    "end": 11.54
+  },
+  {
+    "word": "taking",
+    "start": 11.54,
+    "end": 11.86
+  },
+  {
+    "word": "notes",
+    "start": 11.86,
+    "end": 12.16
+  },
+  {
+    "word": "during",
+    "start": 12.16,
+    "end": 12.54
+  },
+  {
+    "word": "meetings.",
+    "start": 12.54,
+    "end": 12.94
+  },
+  {
+    "word": "It's",
+    "start": 13.6,
+    "end": 13.8
+  },
+  {
+    "word": "amazing",
+    "start": 13.8,
+    "end": 14.1
+  },
+  {
+    "word": "how",
+    "start": 14.1,
+    "end": 14.48
+  },
+  {
+    "word": "it",
+    "start": 14.48,
+    "end": 14.62
+  },
+  {
+    "word": "can",
+    "start": 14.62,
+    "end": 14.74
+  },
+  {
+    "word": "recognise",
+    "start": 14.74,
+    "end": 15.24
+  },
+  {
+    "word": "different",
+    "start": 15.24,
+    "end": 15.68
+  },
+  {
+    "word": "speakers",
+    "start": 15.68,
+    "end": 16.16
+  },
+  {
+    "word": "and",
+    "start": 16.16,
+    "end": 16.8
+  },
+  {
+    "word": "even",
+    "start": 16.8,
+    "end": 17.1
+  },
+  {
+    "word": "add",
+    "start": 17.1,
+    "end": 17.44
+  },
+  {
+    "word": "punctuation.",
+    "start": 17.44,
+    "end": 18.36
+  },
+  {
+    "word": "Yeah,",
+    "start": 18.88,
+    "end": 19.16
+  },
+  {
+    "word": "but",
+    "start": 19.36,
+    "end": 19.52
+  },
+  {
+    "word": "sometimes",
+    "start": 19.52,
+    "end": 20.16
+  },
+  {
+    "word": "noise",
+    "start": 20.16,
+    "end": 20.54
+  },
+  {
+    "word": "can",
+    "start": 20.54,
+    "end": 20.8
+  },
+  {
+    "word": "still",
+    "start": 20.8,
+    "end": 21.1
+  },
+  {
+    "word": "cause",
+    "start": 21.1,
+    "end": 21.44
+  },
+  {
+    "word": "mistakes.",
+    "start": 21.44,
+    "end": 21.94
+  },
+  {
+    "word": "Does",
+    "start": 22.68,
+    "end": 22.9
+  },
+  {
+    "word": "this",
+    "start": 22.9,
+    "end": 23.12
+  },
+  {
+    "word": "system",
+    "start": 23.12,
+    "end": 23.46
+  },
+  {
+    "word": "handle",
+    "start": 23.46,
+    "end": 23.88
+  },
+  {
+    "word": "that",
+    "start": 23.88,
+    "end": 24.12
+  },
+  {
+    "word": "well?",
+    "start": 24.12,
+    "end": 24.42
+  },
+  {
+    "word": "It",
+    "start": 24.42,
+    "end": 25.32
+  },
+  {
+    "word": "does",
+    "start": 25.32,
+    "end": 25.48
+  },
+  {
+    "word": "a",
+    "start": 25.48,
+    "end": 25.62
+  },
+  {
+    "word": "pretty",
+    "start": 25.62,
+    "end": 25.88
+  },
+  {
+    "word": "good",
+    "start": 25.88,
+    "end": 26.08
+  },
+  {
+    "word": "job",
+    "start": 26.08,
+    "end": 26.32
+  },
+  {
+    "word": "filtering",
+    "start": 26.32,
+    "end": 26.8
+  },
+  {
+    "word": "noise,",
+    "start": 26.8,
+    "end": 27.18
+  },
+  {
+    "word": "especially",
+    "start": 27.36,
+    "end": 28.0
+  },
+  {
+    "word": "with",
+    "start": 28.0,
+    "end": 28.28
+  },
+  {
+    "word": "models",
+    "start": 28.28,
+    "end": 28.62
+  },
+  {
+    "word": "that",
+    "start": 28.62,
+    "end": 28.94
+  },
+  {
+    "word": "use",
+    "start": 28.94,
+    "end": 29.22
+  },
+  {
+    "word": "voice",
+    "start": 29.22,
+    "end": 29.54
+  },
+  {
+    "word": "active.",
+    "start": 29.54,
+    "end": 29.9
+  }
+]

audio_tests/generate_transcripts.py

@@ -0,0 +1,57 @@
+#!/usr/bin/env python3
+"""Generate word-level timestamped transcripts using faster-whisper (offline).
+
+Produces one JSON file per audio with: [{word, start, end}, ...]
+"""
+
+import json
+import os
+from faster_whisper import WhisperModel
+
+AUDIO_DIR = os.path.dirname(os.path.abspath(__file__))
+
+FILES = [
+    ("00_00_07_english_1_speaker.wav", "en"),
+    ("00_00_16_french_1_speaker.wav", "fr"),
+    ("00_00_30_english_3_speakers.wav", "en"),
+]
+
+def main():
+    print("Loading faster-whisper model (base, cpu, float32)...")
+    model = WhisperModel("base", device="cpu", compute_type="float32")
+
+    for filename, lang in FILES:
+        audio_path = os.path.join(AUDIO_DIR, filename)
+        out_path = os.path.join(
+            AUDIO_DIR, filename.rsplit(".", 1)[0] + ".transcript.json"
+        )
+
+        print(f"\n{'='*60}")
+        print(f"Transcribing: {filename} (language={lang})")
+        print(f"{'='*60}")
+
+        segments, info = model.transcribe(
+            audio_path, word_timestamps=True, language=lang
+        )
+
+        words = []
+        for segment in segments:
+            if segment.words:
+                for w in segment.words:
+                    words.append({
+                        "word": w.word.strip(),
+                        "start": round(w.start, 3),
+                        "end": round(w.end, 3),
+                    })
+                    print(f"  {w.start:6.2f} - {w.end:6.2f}  {w.word.strip()}")
+
+        with open(out_path, "w", encoding="utf-8") as f:
+            json.dump(words, f, indent=2, ensure_ascii=False)
+
+        print(f"\n  -> {len(words)} words written to {os.path.basename(out_path)}")
+
+    print("\nDone.")
+
+
+if __name__ == "__main__":
+    main()

chrome-extension/README.md

@@ -6,7 +6,7 @@ Capture the audio of your current tab, transcribe diarize and translate it using
 <img src="https://raw.githubusercontent.com/QuentinFuxa/WhisperLiveKit/refs/heads/main/chrome-extension/demo-extension.png" alt="WhisperLiveKit Demo" width="730">
 
 ## Running this extension
-1. Run `python sync_extension.py` to copy frontend files to the `chrome-extension` directory.
+1. Run `python scripts/sync_extension.py` to copy frontend files to the `chrome-extension` directory.
 2. Load the `chrome-extension` directory in Chrome as an unpacked extension.
 
 

compose.yml

@@ -0,0 +1,52 @@
+services:
+  wlk-gpu-sortformer:
+    build:
+      context: .
+      dockerfile: Dockerfile
+      args:
+        EXTRAS: ${GPU_SORTFORMER_EXTRAS:-cu129,diarization-sortformer}
+    image: wlk:gpu-sortformer
+    gpus: all
+    ports:
+      - "8000:8000"
+    volumes:
+      - hf-cache:/root/.cache/huggingface/hub
+      # - ${HF_TKN_FILE:-./token}:/root/.cache/huggingface/token:ro
+    environment:
+      - HF_TOKEN
+    command: ["--model", "medium", "--diarization", "--pcm-input"]
+
+  wlk-gpu-voxtral:
+    build:
+      context: .
+      dockerfile: Dockerfile
+      args:
+        EXTRAS: ${GPU_VOXTRAL_EXTRAS:-cu129,voxtral-hf,translation}
+    image: wlk:gpu-voxtral
+    gpus: all
+    ports:
+      - "8001:8000"
+    volumes:
+      - hf-cache:/root/.cache/huggingface/hub
+      # - ${HF_TKN_FILE:-./token}:/root/.cache/huggingface/token:ro
+    environment:
+      - HF_TOKEN
+    command: ["--backend", "voxtral", "--pcm-input"]
+
+  wlk-cpu:
+    build:
+      context: .
+      dockerfile: Dockerfile.cpu
+      args:
+        EXTRAS: ${CPU_EXTRAS:-cpu,diarization-diart,translation}
+    image: wlk:cpu
+    ports:
+      - "8000:8000"
+    volumes:
+      - hf-cache:/root/.cache/huggingface/hub
+      # - ${HF_TKN_FILE:-./token}:/root/.cache/huggingface/token:ro
+    environment:
+      - HF_TOKEN
+
+volumes:
+  hf-cache:

docs/available_models.md

@@ -1,109 +0,0 @@
-# Available Whisper model sizes:
-
-- tiny.en (english only)
-- tiny
-- base.en (english only)
-- base
-- small.en (english only)
-- small
-- medium.en (english only)
-- medium
-- large-v1
-- large-v2
-- large-v3
-- large-v3-turbo
-
-## How to choose?
-
-### Language Support
-- **English only**: Use `.en` models for better accuracy and faster processing when you only need English transcription
-- **Multilingual**: Do not use `.en` models.
-
-### Resource Constraints
-- **Limited GPU/CPU or need for very low latency**: Choose `small` or smaller models
-  - `tiny`: Fastest, lowest resource usage, acceptable quality for simple audio
-  - `base`: Good balance of speed and accuracy for basic use cases
-  - `small`: Better accuracy while still being resource-efficient
-- **Good resources available**: Use `large` models for best accuracy
-  - `large-v2`: Excellent accuracy, good multilingual support
-  - `large-v3`: Best overall accuracy and language support
-
-### Special Cases
-- **No translation needed**: Use `large-v3-turbo`
-  - Same transcription quality as `large-v2` but significantly faster
-  - **Important**: Does not translate correctly, only transcribes
-
-### Model Comparison Table
-
-| Model | Speed | Accuracy | Multilingual | Translation | Best Use Case |
-|-------|--------|----------|--------------|-------------|---------------|
-| tiny(.en) | Fastest | Basic | Yes/No | Yes/No | Real-time, low resources |
-| base(.en) | Fast | Good | Yes/No | Yes/No | Balanced performance |
-| small(.en) | Medium | Better | Yes/No | Yes/No | Quality on limited hardware |
-| medium(.en) | Slow | High | Yes/No | Yes/No | High quality, moderate resources |
-| large-v2 | Slowest | Excellent | Yes | Yes | Best overall quality |
-| large-v3 | Slowest | Excellent | Yes | Yes | Maximum accuracy |
-| large-v3-turbo | Fast | Excellent | Yes | No | Fast, high-quality transcription |
-
-### Additional Considerations
-
-**Model Performance**:
-- Accuracy improves significantly from tiny to large models
-- English-only models are ~10-15% more accurate for English audio
-- Newer versions (v2, v3) have better punctuation and formatting
-
-**Hardware Requirements**:
-- `tiny`: ~1GB VRAM
-- `base`: ~1GB VRAM  
-- `small`: ~2GB VRAM
-- `medium`: ~5GB VRAM
-- `large`: ~10GB VRAM
-- `large‑v3‑turbo`: ~6GB VRAM
-
-**Audio Quality Impact**:
-- Clean, clear audio: smaller models may suffice
-- Noisy, accented, or technical audio: larger models recommended
-- Phone/low-quality audio: use at least `small` model
-
-### Quick Decision Tree
-1. English only? → Add `.en` to your choice
-2. Limited resources or need speed? → `small` or smaller
-3. Good hardware and want best quality? → `large-v3`
-4. Need fast, high-quality transcription without translation? → `large-v3-turbo`
-5. Need translation capabilities? → `large-v2` or `large-v3` (avoid turbo)
-
-
-_______________________
-
-# Translation Models and Backend
-
-**Language Support**: ~200 languages
-
-## Distilled Model Sizes Available
-
-| Model | Size | Parameters | VRAM (FP16) | VRAM (INT8) | Quality |
-|-------|------|------------|-------------|-------------|---------|
-| 600M | 2.46 GB | 600M | ~1.5GB | ~800MB | Good, understandable |
-| 1.3B | 5.48 GB | 1.3B | ~3GB | ~1.5GB | Better accuracy, context |
-
-**Quality Impact**: 1.3B has ~15-25% better BLEU scores vs 600M across language pairs.
-
-## Backend Performance
-
-| Backend | Speed vs Base | Memory Usage | Quality Loss |
-|---------|---------------|--------------|--------------|
-| CTranslate2 | 6-10x faster | 40-60% less | ~5% BLEU drop |
-| Transformers | Baseline | High | None |
-| Transformers + MPS (on Apple Silicon) | 2x faster | Medium | None |
-
-**Metrics**:
-- CTranslate2: 50-100+ tokens/sec
-- Transformers: 10-30 tokens/sec
-- Apple Silicon with MPS: Up to 2x faster than CTranslate2
-
-## Quick Decision Matrix
-
-**Choose 600M**: Limited resources, close to 0 lag
-**Choose 1.3B**: Quality matters
-**Choose Transformers**: On Apple Silicon
-

docs/default_and_custom_models.md

@@ -0,0 +1,106 @@
+# Models and Model Paths
+
+## Defaults
+
+**Default Whisper Model**: `base`  
+When no model is specified, WhisperLiveKit uses the `base` model, which provides a good balance of speed and accuracy for most use cases.
+
+**Default Model Cache Directory**: `~/.cache/whisper`  
+Models are automatically downloaded from OpenAI's model hub and cached in this directory. You can override this with `--model_cache_dir`.
+
+**Default Translation Model**: `600M` (NLLB-200-distilled)  
+When translation is enabled, the 600M distilled NLLB model is used by default. This provides good quality with minimal resource usage.
+
+**Default Translation Backend**: `transformers`  
+The translation backend defaults to Transformers. On Apple Silicon, this automatically uses MPS acceleration for better performance.
+
+---
+
+
+## Available Whisper model sizes:
+
+| Available Model    | Speed    | Accuracy  | Multilingual | Translation | Hardware Requirements | Best Use Case                   |
+|--------------------|----------|-----------|--------------|-------------|----------------------|----------------------------------|
+| tiny(.en)          | Fastest  | Basic     | Yes/No       | Yes/No      | ~1GB VRAM            | Real-time, low resources         |
+| base(.en)          | Fast     | Good      | Yes/No       | Yes/No      | ~1GB VRAM            | Balanced performance             |
+| small(.en)         | Medium   | Better    | Yes/No       | Yes/No      | ~2GB VRAM            | Quality on limited hardware      |
+| medium(.en)        | Slow     | High      | Yes/No       | Yes/No      | ~5GB VRAM            | High quality, moderate resources |
+| large-v2           | Slowest  | Excellent | Yes          | Yes         | ~10GB VRAM           | Good overall accuracy & language support          |
+| large-v3           | Slowest  | Excellent | Yes          | Yes         | ~10GB VRAM           | Best overall accuracy & language support                |
+| large-v3-turbo     | Fast     | Excellent | Yes          | No          | ~6GB VRAM            | Fast, high-quality transcription |
+
+
+### How to choose?
+
+#### Language Support
+- **English only**: Use `.en` (ex: `base.en`) models for better accuracy and faster processing when you only need English transcription
+- **Multilingual**: Do not use `.en` models.
+      
+#### Special Cases
+- **No translation needed**: Use `large-v3-turbo`
+  - Same transcription quality as `large-v2` but significantly faster
+  - **Important**: Does not translate correctly, only transcribes
+
+### Additional Considerations
+
+**Model Performance**:
+- Accuracy improves significantly from tiny to large models
+- English-only models are ~10-15% more accurate for English audio
+- Newer versions (v2, v3) have better punctuation and formatting
+
+**Audio Quality Impact**:
+- Clean, clear audio: smaller models may suffice
+- Noisy, accented, or technical audio: larger models recommended
+- Phone/low-quality audio: use at least `small` model
+
+_______________________
+
+
+# Custom Models:
+
+The `--model-path` parameter accepts:
+
+## File Path
+- **`.pt` / `.bin` / `.safetensor` formats** Should be openable by pytorch/safetensor.
+
+## Directory Path (recommended)
+Must contain:
+- **`.pt` / `.bin` / `.safetensor` file** (required for decoder)
+
+May optionally contain:
+- **`.bin` file** - faster-whisper model for encoder (requires faster-whisper)
+- **`weights.npz`** or **`weights.safetensors`** - for encoder (requires whisper-mlx)
+
+## Hugging Face Repo ID
+- Provide the repo ID (e.g. `openai/whisper-large-v3`) and WhisperLiveKit will download and cache the snapshot automatically. For gated repos, authenticate via `huggingface-cli login` first.
+
+To improve speed/reduce hallucinations, you may want to use `scripts/determine_alignment_heads.py` to determine the alignment heads to use for your model, and use the `--custom-alignment-heads` to pass them to WLK. If not, alignment heads are set to be all the heads of the last half layer of decoder.
+
+
+_______________________
+
+# Translation Models and Backend
+
+**Language Support**: ~200 languages
+
+## Distilled Model Sizes Available
+
+| Model | Size | Parameters | VRAM (FP16) | VRAM (INT8) | Quality |
+|-------|------|------------|-------------|-------------|---------|
+| 600M | 2.46 GB | 600M | ~1.5GB | ~800MB | Good, understandable |
+| 1.3B | 5.48 GB | 1.3B | ~3GB | ~1.5GB | Better accuracy, context |
+
+**Quality Impact**: 1.3B has ~15-25% better BLEU scores vs 600M across language pairs.
+
+## Backend Performance
+
+| Backend | Speed vs Base | Memory Usage | Quality Loss |
+|---------|---------------|--------------|--------------|
+| CTranslate2 | 6-10x faster | 40-60% less | ~5% BLEU drop |
+| Transformers | Baseline | High | None |
+| Transformers + MPS (on Apple Silicon) | 2x faster | Medium | None |
+
+**Metrics**:
+- CTranslate2: 50-100+ tokens/sec
+- Transformers: 10-30 tokens/sec
+- Apple Silicon with MPS: Up to 2x faster than CTranslate2

docs/models_compatible_formats.md

@@ -1,19 +0,0 @@
-# Model Path Formats
-
-The `--model-path` parameter accepts:
-
-## File Path
-- **`.pt` / `.bin` / `.safetensor` formats** Should be openable by pytorch/safetensor.
-
-## Directory Path (recommended)
-Must contain:
-- **`.pt` / `.bin` / `.safetensor` file** (required for decoder)
-
-May optionally contain:
-- **`.bin` file** - faster-whisper model for encoder (requires faster-whisper)
-- **`weights.npz`** or **`weights.safetensors`** - for encoder (requires whisper-mlx)
-
-## Hugging Face Repo ID
-- Provide the repo ID (e.g. `openai/whisper-large-v3`) and WhisperLiveKit will download and cache the snapshot automatically. For gated repos, authenticate via `huggingface-cli login` first.
-
-To improve speed/reduce allucinations, you may want to use `scripts/determine_alignment_heads.py` to determine the alignment heads to use for your model, and use the `--custom-alignment-heads` to pass them to WLK. If not, alignement heads are set to be all the heads of the last half layer of decoder.

docs/supported_languages.md

@@ -1,6 +1,114 @@
-# Supported Languages
+# Transcription: Supported Language
 
-WhisperLiveKit supports translation into **201 languages** from the FLORES-200 dataset through the NLLB (No Language Left Behind) translation system. 
+WLK supports transcription in the following languages:
+
+| ISO Code | Language Name        |
+|----------|---------------------|
+| en       | English             |
+| zh       | Chinese             |
+| de       | German              |
+| es       | Spanish             |
+| ru       | Russian             |
+| ko       | Korean              |
+| fr       | French              |
+| ja       | Japanese            |
+| pt       | Portuguese          |
+| tr       | Turkish             |
+| pl       | Polish              |
+| ca       | Catalan             |
+| nl       | Dutch               |
+| ar       | Arabic              |
+| sv       | Swedish             |
+| it       | Italian             |
+| id       | Indonesian          |
+| hi       | Hindi               |
+| fi       | Finnish             |
+| vi       | Vietnamese          |
+| he       | Hebrew              |
+| uk       | Ukrainian           |
+| el       | Greek               |
+| ms       | Malay               |
+| cs       | Czech               |
+| ro       | Romanian            |
+| da       | Danish              |
+| hu       | Hungarian           |
+| ta       | Tamil               |
+| no       | Norwegian           |
+| th       | Thai                |
+| ur       | Urdu                |
+| hr       | Croatian            |
+| bg       | Bulgarian           |
+| lt       | Lithuanian          |
+| la       | Latin               |
+| mi       | Maori               |
+| ml       | Malayalam           |
+| cy       | Welsh               |
+| sk       | Slovak              |
+| te       | Telugu              |
+| fa       | Persian             |
+| lv       | Latvian             |
+| bn       | Bengali             |
+| sr       | Serbian             |
+| az       | Azerbaijani         |
+| sl       | Slovenian           |
+| kn       | Kannada             |
+| et       | Estonian            |
+| mk       | Macedonian          |
+| br       | Breton              |
+| eu       | Basque              |
+| is       | Icelandic           |
+| hy       | Armenian            |
+| ne       | Nepali              |
+| mn       | Mongolian           |
+| bs       | Bosnian             |
+| kk       | Kazakh              |
+| sq       | Albanian            |
+| sw       | Swahili             |
+| gl       | Galician            |
+| mr       | Marathi             |
+| pa       | Punjabi             |
+| si       | Sinhala             |
+| km       | Khmer               |
+| sn       | Shona               |
+| yo       | Yoruba              |
+| so       | Somali              |
+| af       | Afrikaans           |
+| oc       | Occitan             |
+| ka       | Georgian            |
+| be       | Belarusian          |
+| tg       | Tajik               |
+| sd       | Sindhi              |
+| gu       | Gujarati            |
+| am       | Amharic             |
+| yi       | Yiddish             |
+| lo       | Lao                 |
+| uz       | Uzbek               |
+| fo       | Faroese             |
+| ht       | Haitian Creole      |
+| ps       | Pashto              |
+| tk       | Turkmen             |
+| nn       | Nynorsk             |
+| mt       | Maltese             |
+| sa       | Sanskrit            |
+| lb       | Luxembourgish       |
+| my       | Myanmar             |
+| bo       | Tibetan             |
+| tl       | Tagalog             |
+| mg       | Malagasy            |
+| as       | Assamese            |
+| tt       | Tatar               |
+| haw      | Hawaiian            |
+| ln       | Lingala             |
+| ha       | Hausa               |
+| ba       | Bashkir             |
+| jw       | Javanese            |
+| su       | Sundanese           |
+| yue      | Cantonese           |
+
+
+# Translation: Supported Languages 
+
+WLK supports translation into **201 languages** from the FLORES-200 dataset through the [NLLW](https://github.com/QuentinFuxa/NoLanguageLeftWaiting) translation system. 
 
 ## How to Specify Languages
 

docs/technical_integration.md

@@ -40,4 +40,4 @@ This document introduce how to reuse the core components when you do **not** wan
 3. Call `create_tasks()` to get the async generator, `process_audio()` with incoming bytes, and ensure `cleanup()` runs when the client disconnects.
 
 
-If you prefer to send compressed audio, instantiate `AudioProcessor(pcm_input=False)` and pipe encoded chunks through `FFmpegManager` transparently—just ensure `ffmpeg` is available or be ready to handle the `"ffmpeg_not_found"` error in the streamed `FrontData`.
+If you prefer to send compressed audio, instantiate `AudioProcessor(pcm_input=False)` and pipe encoded chunks through `FFmpegManager` transparently. Just ensure `ffmpeg` is available.

docs/troubleshooting.md

@@ -0,0 +1,140 @@
+# Troubleshooting
+
+
+## GPU drivers & cuDNN visibility
+
+### Linux error: `Unable to load libcudnn_ops.so* / cudnnCreateTensorDescriptor`
+> Reported in issue #271 (Arch/CachyOS)
+
+`faster-whisper` (used for the SimulStreaming encoder) dynamically loads cuDNN.  
+If the runtime cannot find `libcudnn_*`, verify that CUDA and cuDNN match the PyTorch build you installed:
+
+1. **Install CUDA + cuDNN** (Arch/CachyOS example):
+   ```bash
+   sudo pacman -S cuda cudnn
+   sudo ldconfig
+   ```
+2. **Make sure the shared objects are visible**:
+   ```bash
+   ls /usr/lib/libcudnn*
+   ```
+3. **Check what CUDA version PyTorch expects** and match that with the driver you installed:
+   ```bash
+   python - <<'EOF'
+   import torch
+   print(torch.version.cuda)
+   EOF
+   nvcc --version
+   ```
+4. If you installed CUDA in a non-default location, export `CUDA_HOME` and add `$CUDA_HOME/lib64` to `LD_LIBRARY_PATH`.
+
+Once the CUDA/cuDNN versions match, `whisperlivekit-server` starts normally.
+
+### Windows error: `Could not locate cudnn_ops64_9.dll`
+> Reported in issue #286 (Conda on Windows)
+
+PyTorch bundles cuDNN DLLs inside your environment (`<env>\Lib\site-packages\torch\lib`).  
+When `ctranslate2` or `faster-whisper` cannot find `cudnn_ops64_9.dll`:
+
+1. Locate the DLL shipped with PyTorch, e.g.
+   ```
+   E:\conda\envs\WhisperLiveKit\Lib\site-packages\torch\lib\cudnn_ops64_9.dll
+   ```
+2. Add that directory to your `PATH` **or** copy the `cudnn_*64_9.dll` files into a directory that is already on `PATH` (such as the environment's `Scripts/` folder).
+3. Restart the shell before launching `wlk`.
+
+Installing NVIDIA's standalone cuDNN 9.x and pointing `PATH`/`CUDNN_PATH` to it works as well, but is usually not required.
+
+---
+
+## PyTorch / CTranslate2 GPU builds
+
+### `Torch not compiled with CUDA enabled`
+> Reported in issue #284
+
+If `torch.zeros(1).cuda()` raises that assertion it means you installed a CPU-only wheel.  
+Install the GPU-enabled wheels that match your CUDA toolkit:
+
+```bash
+pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130
+```
+
+Replace `cu130` with the CUDA version supported by your driver (see [PyTorch install selector](https://pytorch.org/get-started/locally/)).  
+Validate with:
+
+```python
+import torch
+print(torch.cuda.is_available(), torch.cuda.get_device_name())
+```
+
+### `CTranslate2 device count: 0` or `Could not infer dtype of ctranslate2._ext.StorageView`
+> Follow-up in issue #284
+
+`ctranslate2` publishes separate CPU and CUDA wheels. The default `pip install ctranslate2` brings the CPU build, which makes WhisperLiveKit fall back to CPU tensors and leads to the dtype error above.
+
+1. Uninstall the CPU build: `pip uninstall -y ctranslate2`.
+2. Install the CUDA wheel that matches your toolkit (example for CUDA 13.0):
+   ```bash
+   pip install ctranslate2==4.5.0 -f https://opennmt.net/ctranslate2/whl/cu130
+   ```
+   (See the [CTranslate2 installation table](https://opennmt.net/CTranslate2/installation.html) for other CUDA versions.)
+3. Verify:
+   ```python
+   import ctranslate2
+   print("CUDA devices:", ctranslate2.get_cuda_device_count())
+   print("CUDA compute types:", ctranslate2.get_supported_compute_types("cuda", 0))
+   ```
+
+**Note for aarch64 systems (e.g., NVIDIA DGX Spark):** Pre-built CUDA wheels may not be available for all CUDA versions on ARM architectures. If the wheel installation fails, you may need to compile CTranslate2 from source with CUDA support enabled.
+
+If you intentionally want CPU inference, run `wlk --backend whisper` to avoid mixing CPU-only CTranslate2 with a GPU Torch build.
+
+---
+
+## Hopper / Blackwell (`sm_121a`) systems
+> Reported in issues #276 and #284 (NVIDIA DGX Spark)
+
+CUDA 12.1a GPUs (e.g., NVIDIA GB10 on DGX Spark) ship before some toolchains know about the architecture ID, so Triton/PTXAS need manual configuration.
+
+### Error: `ptxas fatal : Value 'sm_121a' is not defined for option 'gpu-name'`
+
+If you encounter this error after compiling CTranslate2 from source on aarch64 systems, Triton's bundled `ptxas` may not support the `sm_121a` architecture. The solution is to replace Triton's `ptxas` with the system's CUDA `ptxas`:
+
+```bash
+# Find your Python environment's Triton directory
+python -c "import triton; import os; print(os.path.dirname(triton.__file__))"
+
+# Copy the system ptxas to Triton's backend directory
+# Replace <triton_path> with the output above
+cp /usr/local/cuda/bin/ptxas <triton_path>/backends/nvidia/bin/ptxas
+```
+
+For example, in a virtual environment:
+```bash
+cp /usr/local/cuda/bin/ptxas ~/wlk/lib/python3.12/site-packages/triton/backends/nvidia/bin/ptxas
+```
+
+**Note:** On DGX Spark systems, CUDA is typically already in `PATH` (`/usr/local/cuda/bin`), so explicit `CUDA_HOME` and `PATH` exports may not be necessary. Verify with `which ptxas` before copying.
+
+### Alternative: Environment variable approach
+
+If the above doesn't work, you can try setting environment variables (though this may not resolve the `sm_121a` issue on all systems):
+
+```bash
+export CUDA_HOME="/usr/local/cuda-13.0"
+export PATH="$CUDA_HOME/bin:$PATH"
+export LD_LIBRARY_PATH="$CUDA_HOME/lib64:$LD_LIBRARY_PATH"
+
+# Tell Triton where the new ptxas lives
+export TRITON_PTXAS_PATH="$CUDA_HOME/bin/ptxas"
+
+# Force PyTorch to JIT kernels for all needed architectures
+export TORCH_CUDA_ARCH_LIST="8.0 9.0 10.0 12.0 12.1a"
+```
+
+After applying the fix, restart `wlk`. Incoming streams will now compile kernels targeting `sm_121a` without crashing.
+
+---
+
+Need help with another recurring issue? Open a GitHub discussion or PR and reference this document so we can keep it current.
+

pyproject.toml

@@ -4,27 +4,21 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "whisperlivekit"
-version = "0.2.15"
+version = "0.2.19"
 description = "Real-time speech-to-text with speaker diarization using Whisper"
 readme = "README.md"
-authors = [
-    { name = "Quentin Fuxa" }
-]
+authors = [{ name = "Quentin Fuxa" }]
 license = { file = "LICENSE" }
-requires-python = ">=3.9"
+requires-python = ">=3.11, <3.14"
 classifiers = [
     "Development Status :: 4 - Beta",
     "Intended Audience :: Developers",
     "License :: OSI Approved :: MIT License",
-    "Programming Language :: Python :: 3.9",
-    "Programming Language :: Python :: 3.10",
     "Programming Language :: Python :: 3.11",
     "Programming Language :: Python :: 3.12",
     "Programming Language :: Python :: 3.13",
-    "Programming Language :: Python :: 3.14",
-    "Programming Language :: Python :: 3.15",
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
-    "Topic :: Multimedia :: Sound/Audio :: Speech"
+    "Topic :: Multimedia :: Sound/Audio :: Speech",
 ]
 dependencies = [
     "fastapi",
@@ -32,17 +26,91 @@ dependencies = [
     "soundfile",
     "uvicorn",
     "websockets",
-    "torchaudio>=2.0.0",
-    "torch>=2.0.0",
     "huggingface-hub>=0.25.0",
+    "faster-whisper>=1.2.0",
+    "torch>=2.0.0",
+    "torchaudio>=2.0.0",
     "tqdm",
     "tiktoken",
-    'triton>=2.0.0; platform_machine == "x86_64" and (sys_platform == "linux" or sys_platform == "linux2")'
 ]
 
 [project.optional-dependencies]
+test = ["pytest>=7.0", "pytest-asyncio>=0.21"]
 translation = ["nllw"]
 sentence_tokenizer = ["mosestokenizer", "wtpsplit"]
+mlx-whisper = [
+    'mlx>=0.11.0; sys_platform == "darwin" and platform_machine == "arm64"',
+    'mlx-whisper>=0.4.0; sys_platform == "darwin" and platform_machine == "arm64"',
+]
+voxtral-mlx = [
+    'mlx>=0.11.0; sys_platform == "darwin" and platform_machine == "arm64"',
+    'mlx-whisper>=0.4.0; sys_platform == "darwin" and platform_machine == "arm64"',
+    "mistral-common[audio]",
+]
+voxtral-hf = [
+    "transformers>=5.2.0; python_version >= '3.10'",
+    "mistral-common[audio]",
+    "accelerate>=0.12",
+]
+cpu = ["torch>=2.0.0", "torchaudio>=2.0.0"]
+cu129 = [
+    "torch>=2.0.0",
+    "torchaudio>=2.0.0",
+    'triton>=2.0.0; platform_machine == "x86_64" and (sys_platform == "linux" or sys_platform == "linux2")',
+]
+diarization-sortformer = [
+    "nemo-toolkit[asr]>2.4; python_version >= '3.10' and python_version < '3.13'",
+]
+diarization-diart = [
+    "diart",
+    "torch<2.9.0",
+    "torchaudio<2.9.0",
+    "torchvision<0.24.0",
+]
+
+[dependency-groups]
+dev = ["rich>=14.3.3"]
+
+[tool.uv]
+conflicts = [
+    [
+        { extra = "cpu" },
+        { extra = "cu129" },
+    ],
+    [
+        { extra = "diarization-diart" },
+        { extra = "cu129" },
+    ],
+    [
+        { extra = "voxtral-hf" },
+        { extra = "diarization-sortformer" },
+    ],
+]
+
+[tool.uv.sources]
+torch = [
+    { index = "pytorch-cpu", extra = "cpu", marker = "platform_system != 'Darwin'" },
+    { index = "pytorch-cpu", extra = "diarization-diart", marker = "platform_system != 'Darwin'" },
+    { index = "pytorch-cu129", extra = "cu129", marker = "platform_system == 'Linux' and platform_machine == 'x86_64'" },
+]
+torchaudio = [
+    { index = "pytorch-cpu", extra = "cpu", marker = "platform_system != 'Darwin'" },
+    { index = "pytorch-cpu", extra = "diarization-diart", marker = "platform_system != 'Darwin'" },
+    { index = "pytorch-cu129", extra = "cu129", marker = "platform_system == 'Linux' and platform_machine == 'x86_64'" },
+]
+torchvision = [
+    { index = "pytorch-cpu", extra = "diarization-diart", marker = "platform_system != 'Darwin'" },
+]
+
+[[tool.uv.index]]
+name = "pytorch-cpu"
+url = "https://download.pytorch.org/whl/cpu"
+explicit = true
+
+[[tool.uv.index]]
+name = "pytorch-cu129"
+url = "https://download.pytorch.org/whl/cu129"
+explicit = true
 
 [project.urls]
 Homepage = "https://github.com/QuentinFuxa/WhisperLiveKit"
@@ -56,15 +124,18 @@ packages = [
     "whisperlivekit",
     "whisperlivekit.diarization",
     "whisperlivekit.simul_whisper",
+    "whisperlivekit.simul_whisper.mlx",
     "whisperlivekit.whisper",
     "whisperlivekit.whisper.assets",
     "whisperlivekit.whisper.normalizers",
     "whisperlivekit.web",
     "whisperlivekit.local_agreement",
-    "whisperlivekit.vad_models"
+    "whisperlivekit.voxtral_mlx",
+    "whisperlivekit.silero_vad_models",
 ]
 
 [tool.setuptools.package-data]
 whisperlivekit = ["web/*.html", "web/*.css", "web/*.js", "web/src/*.svg"]
 "whisperlivekit.whisper.assets" = ["*.tiktoken", "*.npz"]
-"whisperlivekit.vad_models" = ["*.jit", "*.onnx"]
+"whisperlivekit.whisper.normalizers" = ["*.json"]
+"whisperlivekit.silero_vad_models" = ["*.jit", "*.onnx"]

run_benchmark.py

@@ -0,0 +1,291 @@
+#!/usr/bin/env python3
+"""
+Comprehensive benchmark runner for WhisperLiveKit.
+
+Tests all available backend+policy combinations across multiple audio files,
+model sizes, and VAC on/off configurations. Outputs structured JSON that
+is consumed by the report generator.
+
+Usage:
+    python run_benchmark.py                    # full benchmark
+    python run_benchmark.py --quick            # subset (tiny models, fewer combos)
+    python run_benchmark.py --json results.json # custom output path
+"""
+
+import argparse
+import asyncio
+import gc
+import json
+import logging
+import platform
+import subprocess
+import sys
+import time
+from dataclasses import asdict
+from pathlib import Path
+
+logging.basicConfig(level=logging.WARNING, format="%(asctime)s %(levelname)s %(name)s: %(message)s")
+logger = logging.getLogger("benchmark")
+logger.setLevel(logging.INFO)
+
+# Re-use harness functions
+sys.path.insert(0, str(Path(__file__).parent))
+from test_backend_offline import (
+    AUDIO_TESTS_DIR,
+    SAMPLE_RATE,
+    TestResult,
+    create_engine,
+    discover_audio_files,
+    download_sample_audio,
+    load_audio,
+    run_test,
+)
+
+CACHE_DIR = Path(__file__).parent / ".test_cache"
+
+
+def get_system_info() -> dict:
+    """Collect system metadata for the report."""
+    info = {
+        "platform": platform.platform(),
+        "machine": platform.machine(),
+        "processor": platform.processor(),
+        "python_version": platform.python_version(),
+    }
+
+    # macOS: get chip info
+    try:
+        chip = subprocess.check_output(
+            ["sysctl", "-n", "machdep.cpu.brand_string"], text=True
+        ).strip()
+        info["cpu"] = chip
+    except Exception:
+        info["cpu"] = platform.processor()
+
+    # RAM
+    try:
+        mem_bytes = int(
+            subprocess.check_output(["sysctl", "-n", "hw.memsize"], text=True).strip()
+        )
+        info["ram_gb"] = round(mem_bytes / (1024**3))
+    except Exception:
+        info["ram_gb"] = None
+
+    # Backend versions
+    versions = {}
+    try:
+        import faster_whisper
+        versions["faster-whisper"] = faster_whisper.__version__
+    except ImportError:
+        pass
+    try:
+        import mlx_whisper  # noqa: F401
+        versions["mlx-whisper"] = "installed"
+    except ImportError:
+        pass
+    try:
+        import mlx.core as mx
+        versions["mlx"] = mx.__version__
+    except ImportError:
+        pass
+    try:
+        import transformers
+        versions["transformers"] = transformers.__version__
+    except ImportError:
+        pass
+    try:
+        import torch
+        versions["torch"] = torch.__version__
+    except ImportError:
+        pass
+
+    info["backend_versions"] = versions
+    return info
+
+
+def detect_combos(quick: bool = False) -> list:
+    """Build list of (backend, policy, model_size) combos to test."""
+    combos = []
+
+    # Model sizes to test
+    model_sizes = ["tiny", "base", "small"] if not quick else ["tiny", "base"]
+
+    # faster-whisper
+    try:
+        import faster_whisper  # noqa: F401
+        for model in model_sizes:
+            combos.append({"backend": "faster-whisper", "policy": "localagreement", "model": model})
+            combos.append({"backend": "faster-whisper", "policy": "simulstreaming", "model": model})
+    except ImportError:
+        pass
+
+    # mlx-whisper
+    try:
+        import mlx_whisper  # noqa: F401
+        for model in model_sizes:
+            combos.append({"backend": "mlx-whisper", "policy": "localagreement", "model": model})
+            combos.append({"backend": "mlx-whisper", "policy": "simulstreaming", "model": model})
+    except ImportError:
+        pass
+
+    # voxtral-mlx (single model, single policy)
+    try:
+        from whisperlivekit.voxtral_mlx import VoxtralMLXModel  # noqa: F401
+        combos.append({"backend": "voxtral-mlx", "policy": "voxtral", "model": ""})
+    except ImportError:
+        pass
+
+    # voxtral HF (single model, single policy)
+    try:
+        from transformers import AutoModelForSpeechSeq2Seq  # noqa: F401
+        combos.append({"backend": "voxtral", "policy": "voxtral", "model": ""})
+    except ImportError:
+        pass
+
+    return combos
+
+
+def collect_audio_files() -> list:
+    """Collect all benchmark audio files."""
+    files = []
+
+    # audio_tests/ directory
+    if AUDIO_TESTS_DIR.is_dir():
+        files.extend(discover_audio_files(str(AUDIO_TESTS_DIR)))
+
+    # JFK sample
+    jfk = CACHE_DIR / "jfk.wav"
+    if not jfk.exists():
+        jfk = download_sample_audio()
+    if jfk.exists():
+        files.append(jfk)
+
+    return files
+
+
+async def run_single_combo(
+    combo: dict, audio_files: list, vac: bool, lan: str, max_duration: float,
+) -> list:
+    """Run one backend+policy+model combo across all audio files."""
+    backend = combo["backend"]
+    policy = combo["policy"]
+    model = combo["model"]
+
+    results = []
+    try:
+        engine = create_engine(
+            backend=backend,
+            model_size=model,
+            lan=lan,
+            vac=vac,
+            policy=policy,
+        )
+
+        # Quiet noisy loggers
+        for mod in (
+            "whisperlivekit.audio_processor",
+            "whisperlivekit.simul_whisper",
+            "whisperlivekit.tokens_alignment",
+            "whisperlivekit.simul_whisper.align_att_base",
+            "whisperlivekit.simul_whisper.simul_whisper",
+        ):
+            logging.getLogger(mod).setLevel(logging.WARNING)
+
+        for audio_path in audio_files:
+            duration = len(load_audio(str(audio_path))) / SAMPLE_RATE
+            if duration > max_duration:
+                logger.info(f"  Skipping {audio_path.name} ({duration:.0f}s > {max_duration:.0f}s)")
+                continue
+
+            file_lan = lan
+            if "french" in audio_path.name.lower() and lan == "en":
+                file_lan = "fr"
+
+            audio = load_audio(str(audio_path))
+            result = await run_test(
+                engine, audio, chunk_ms=100, realtime=False,
+                audio_file=audio_path.name, backend=backend,
+                policy=policy, lan=file_lan,
+            )
+            # Tag with extra metadata
+            result_dict = asdict(result)
+            result_dict["model_size"] = model
+            result_dict["vac"] = vac
+            results.append(result_dict)
+
+    except Exception as e:
+        logger.error(f"  FAILED: {e}")
+        import traceback
+        traceback.print_exc()
+
+    return results
+
+
+async def run_full_benchmark(combos, audio_files, max_duration=60.0):
+    """Run all combos with VAC on and off."""
+    all_results = []
+    total = len(combos) * 2  # x2 for VAC on/off
+    idx = 0
+
+    for combo in combos:
+        for vac in [True, False]:
+            idx += 1
+            vac_str = "VAC=on" if vac else "VAC=off"
+            desc = f"{combo['backend']} / {combo['policy']}"
+            if combo["model"]:
+                desc += f" / {combo['model']}"
+            desc += f" / {vac_str}"
+
+            print(f"\n{'='*70}")
+            print(f"[{idx}/{total}] {desc}")
+            print(f"{'='*70}")
+
+            results = await run_single_combo(
+                combo, audio_files, vac=vac, lan="en", max_duration=max_duration,
+            )
+            all_results.extend(results)
+
+            # Free memory between combos
+            gc.collect()
+
+    return all_results
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Run comprehensive WhisperLiveKit benchmark")
+    parser.add_argument("--quick", action="store_true", help="Quick mode: fewer models and combos")
+    parser.add_argument("--json", default="benchmark_results.json", dest="json_output", help="Output JSON path")
+    parser.add_argument("--max-duration", type=float, default=60.0, help="Max audio duration in seconds")
+    args = parser.parse_args()
+
+    system_info = get_system_info()
+    combos = detect_combos(quick=args.quick)
+    audio_files = collect_audio_files()
+
+    print(f"System: {system_info.get('cpu', 'unknown')}, {system_info.get('ram_gb', '?')}GB RAM")
+    print(f"Backends: {list(system_info['backend_versions'].keys())}")
+    print(f"Combos to test: {len(combos)} x 2 (VAC on/off) = {len(combos)*2}")
+    print(f"Audio files: {[f.name for f in audio_files]}")
+    print()
+
+    t0 = time.time()
+    all_results = asyncio.run(
+        run_full_benchmark(combos, audio_files, max_duration=args.max_duration)
+    )
+    total_time = time.time() - t0
+
+    output = {
+        "system_info": system_info,
+        "benchmark_date": time.strftime("%Y-%m-%d %H:%M"),
+        "total_benchmark_time_s": round(total_time, 1),
+        "n_combos": len(combos) * 2,
+        "n_audio_files": len(audio_files),
+        "results": all_results,
+    }
+
+    Path(args.json_output).write_text(json.dumps(output, indent=2, ensure_ascii=False))
+    print(f"\nBenchmark complete in {total_time:.0f}s. Results: {args.json_output}")
+
+
+if __name__ == "__main__":
+    main()

scripts/convert_hf_whisper.py

@@ -14,10 +14,10 @@ from typing import Dict, Tuple
 
 import torch
 
+from whisperlivekit.whisper import _convert_hf_state_dict
 from whisperlivekit.whisper.audio import HOP_LENGTH, SAMPLE_RATE
 from whisperlivekit.whisper.model import ModelDimensions
 from whisperlivekit.whisper.utils import exact_div
-from whisperlivekit.whisper import _convert_hf_state_dict
 
 
 def _load_state_dict(repo_path: Path) -> Dict[str, torch.Tensor]:

scripts/determine_alignment_heads.py

@@ -5,16 +5,18 @@ import argparse
 import base64
 import gzip
 import io
+import math
 import pathlib
 import sys
-import math
 from typing import List, Optional, Sequence, Tuple, Union
 
-import numpy as np
-import torch
-from datasets import Audio as DatasetAudio, load_dataset
-import soundfile as sf
 import matplotlib.pyplot as plt
+import numpy as np
+import soundfile as sf
+import torch
+from datasets import Audio as DatasetAudio
+from datasets import load_dataset
+
 REPO_ROOT = pathlib.Path(__file__).resolve().parents[1]
 WHISPER_ROOT = REPO_ROOT / "whisper"
 

scripts/python_support_matrix.py

@@ -0,0 +1,580 @@
+#!/usr/bin/env python3
+"""Offline Python support matrix runner for WhisperLiveKit."""
+
+from __future__ import annotations
+
+import argparse
+import os
+import shlex
+import shutil
+import subprocess
+import sys
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Literal
+
+try:
+    from rich.console import Console
+    from rich.table import Table
+
+    HAS_RICH = True
+except Exception:
+    HAS_RICH = False
+
+SAMPLE_URL = (
+    "https://github.com/pyannote/pyannote-audio/raw/develop/tutorials/assets/sample.wav"
+)
+SAMPLE_PATH = Path("audio_tests/support-matrix-sample.wav")
+DEFAULT_LOGS_DIR = Path("outputs/python-matrix/logs")
+PYTHON_VERSIONS = ("3.11", "3.12", "3.13")
+CONSOLE = Console() if HAS_RICH else None
+
+
+@dataclass(frozen=True)
+class MatrixRow:
+    row_id: str
+    extras: tuple[str, ...]
+    backend: str
+    policy: str
+    diarization_backend: str
+    requires_gpu: bool = False
+
+
+CASES = (
+    MatrixRow(
+        row_id="fw-diart-cpu",
+        extras=("test", "cpu", "diarization-diart"),
+        backend="faster-whisper",
+        policy="simulstreaming",
+        diarization_backend="diart",
+    ),
+    MatrixRow(
+        row_id="fw-sortformer-cpu",
+        extras=("test", "cpu", "diarization-sortformer"),
+        backend="faster-whisper",
+        policy="simulstreaming",
+        diarization_backend="sortformer",
+    ),
+    MatrixRow(
+        row_id="fw-sortformer-gpu",
+        extras=("test", "cu129", "diarization-sortformer"),
+        backend="faster-whisper",
+        policy="simulstreaming",
+        diarization_backend="sortformer",
+        requires_gpu=True,
+    ),
+    MatrixRow(
+        row_id="voxtral-diart-cpu",
+        extras=("test", "cpu", "voxtral-hf", "diarization-diart"),
+        backend="voxtral",
+        policy="voxtral",
+        diarization_backend="diart",
+    ),
+)
+
+EXPECTED_FAILURE_CASES = {
+    ("3.11", "voxtral-diart-cpu"): "known_unstable_voxtral_diart_cpu",
+    ("3.12", "voxtral-diart-cpu"): "known_unstable_voxtral_diart_cpu",
+}
+UNSUPPORTED_CASES = {
+    ("3.13", "fw-sortformer-cpu"): "unsupported_py313_sortformer_protobuf",
+    ("3.13", "fw-sortformer-gpu"): "unsupported_py313_sortformer_protobuf",
+}
+
+
+@dataclass(frozen=True)
+class CaseResult:
+    python_version: str
+    row_id: str
+    status: Literal["PASS", "FAIL", "N/A"]
+    reason: str
+    duration_sec: float
+    hint: str = ""
+    log_path: str = ""
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Minimal WhisperLiveKit offline support matrix"
+    )
+    parser.add_argument(
+        "--timeout-sec",
+        type=int,
+        default=300,
+        help="Per-case timeout in seconds (default: 300)",
+    )
+    parser.add_argument(
+        "--logs-dir",
+        default=str(DEFAULT_LOGS_DIR),
+        help="Directory where per-case logs are written (default: outputs/python-matrix/logs)",
+    )
+    return parser.parse_args()
+
+
+def safe_slug(text: str) -> str:
+    return text.replace("=", "-").replace("|", "__").replace("/", "-").replace(" ", "-")
+
+
+def status_style(status: str) -> str:
+    if status == "PASS":
+        return "green"
+    if status == "FAIL":
+        return "bold red"
+    if status == "N/A":
+        return "yellow"
+    return "white"
+
+
+def print_line(message: str, style: str | None = None) -> None:
+    if CONSOLE is None:
+        print(message)
+        return
+    if style:
+        CONSOLE.print(message, style=style, highlight=False)
+    else:
+        CONSOLE.print(message, highlight=False)
+
+
+def tail_text(text: str | None, max_chars: int = 220) -> str:
+    if not text:
+        return ""
+    normalized = " ".join(text.split())
+    if len(normalized) <= max_chars:
+        return normalized
+    return normalized[-max_chars:]
+
+
+def run_command(
+    cmd: list[str],
+    cwd: Path,
+    env: dict[str, str],
+    timeout: int | None = None,
+    log_path: Path | None = None,
+    log_section: str | None = None,
+) -> subprocess.CompletedProcess[str]:
+    def _append_log(
+        *,
+        command: list[str],
+        section: str,
+        returncode: int | None,
+        stdout: str | None,
+        stderr: str | None,
+        timed_out: bool = False,
+    ) -> None:
+        if log_path is None:
+            return
+        log_path.parent.mkdir(parents=True, exist_ok=True)
+        with log_path.open("a", encoding="utf-8") as f:
+            f.write(f"\n=== {section} ===\n")
+            f.write(f"$ {shlex.join(command)}\n")
+            if timed_out:
+                f.write("status: timeout\n")
+            else:
+                f.write(f"status: exit_code={returncode}\n")
+            if stdout:
+                f.write("--- stdout ---\n")
+                f.write(stdout)
+                if not stdout.endswith("\n"):
+                    f.write("\n")
+            if stderr:
+                f.write("--- stderr ---\n")
+                f.write(stderr)
+                if not stderr.endswith("\n"):
+                    f.write("\n")
+
+    section = log_section or "command"
+    try:
+        proc = subprocess.run(
+            cmd,
+            cwd=str(cwd),
+            env=env,
+            text=True,
+            capture_output=True,
+            check=False,
+            timeout=timeout,
+        )
+    except subprocess.TimeoutExpired as exc:
+        _append_log(
+            command=cmd,
+            section=section,
+            returncode=None,
+            stdout=exc.stdout if isinstance(exc.stdout, str) else None,
+            stderr=exc.stderr if isinstance(exc.stderr, str) else None,
+            timed_out=True,
+        )
+        raise
+
+    _append_log(
+        command=cmd,
+        section=section,
+        returncode=proc.returncode,
+        stdout=proc.stdout,
+        stderr=proc.stderr,
+    )
+    return proc
+
+
+def detect_gpu_available() -> bool:
+    try:
+        proc = subprocess.run(
+            ["nvidia-smi", "-L"],
+            text=True,
+            capture_output=True,
+            check=False,
+            timeout=10,
+        )
+    except (FileNotFoundError, subprocess.TimeoutExpired):
+        return False
+    return proc.returncode == 0
+
+
+def download_sample(repo_root: Path) -> Path:
+    target = repo_root / SAMPLE_PATH
+    target.parent.mkdir(parents=True, exist_ok=True)
+    cmd = [
+        "curl",
+        "--fail",
+        "--location",
+        "--silent",
+        "--show-error",
+        SAMPLE_URL,
+        "--output",
+        str(target),
+    ]
+    proc = run_command(cmd, cwd=repo_root, env=os.environ.copy())
+    if proc.returncode != 0:
+        hint = tail_text(proc.stderr or proc.stdout)
+        raise RuntimeError(f"sample_download_failed: {hint}")
+    return target
+
+
+def sync_case_environment(
+    repo_root: Path,
+    python_version: str,
+    row: MatrixRow,
+    env_dir: Path,
+    log_path: Path,
+) -> tuple[bool, str]:
+    cmd = ["uv", "sync", "--python", python_version, "--no-dev"]
+    for extra in row.extras:
+        cmd.extend(["--extra", extra])
+    env = os.environ.copy()
+    env["UV_PROJECT_ENVIRONMENT"] = str(env_dir)
+    proc = run_command(
+        cmd,
+        cwd=repo_root,
+        env=env,
+        log_path=log_path,
+        log_section="sync",
+    )
+    if proc.returncode != 0:
+        return False, tail_text(proc.stderr or proc.stdout)
+    return True, ""
+
+
+def apply_expected_failure_policy(result: CaseResult) -> CaseResult:
+    expected_reason = EXPECTED_FAILURE_CASES.get((result.python_version, result.row_id))
+    if result.status != "FAIL" or not expected_reason:
+        return result
+    override_hint = result.hint
+    if result.reason:
+        override_hint = (
+            f"expected_failure_override original_reason={result.reason}; {override_hint}"
+            if override_hint
+            else f"expected_failure_override original_reason={result.reason}"
+        )
+    return CaseResult(
+        python_version=result.python_version,
+        row_id=result.row_id,
+        status="N/A",
+        reason=expected_reason,
+        duration_sec=result.duration_sec,
+        hint=override_hint,
+        log_path=result.log_path,
+    )
+
+
+def build_offline_command(
+    python_version: str,
+    row: MatrixRow,
+    sample_audio: Path,
+    timeout_sec: int,
+) -> tuple[list[str], int | None]:
+    base_cmd = [
+        "uv",
+        "run",
+        "--python",
+        python_version,
+        "--no-sync",
+        "python",
+        "test_backend_offline.py",
+        "--backend",
+        row.backend,
+        "--policy",
+        row.policy,
+        "--audio",
+        str(sample_audio),
+        "--model",
+        "tiny",
+        "--diarization",
+        "--diarization-backend",
+        row.diarization_backend,
+        "--lan",
+        "en",
+        "--no-realtime",
+    ]
+    if shutil.which("timeout"):
+        return ["timeout", str(timeout_sec), *base_cmd], None
+    return base_cmd, timeout_sec
+
+
+def run_case(
+    repo_root: Path,
+    python_version: str,
+    row: MatrixRow,
+    sample_audio: Path,
+    timeout_sec: int,
+    gpu_available: bool,
+    logs_dir: Path,
+) -> CaseResult:
+    start = time.monotonic()
+    case_slug = safe_slug(f"py{python_version}-{row.row_id}")
+    log_path = logs_dir / f"run-{case_slug}.log"
+    log_path.parent.mkdir(parents=True, exist_ok=True)
+    log_path.write_text("", encoding="utf-8")
+
+    unsupported_reason = UNSUPPORTED_CASES.get((python_version, row.row_id))
+    if unsupported_reason:
+        log_path.write_text(
+            f"[matrix] precheck_short_circuit status=N/A reason={unsupported_reason}\n",
+            encoding="utf-8",
+        )
+        return CaseResult(
+            python_version=python_version,
+            row_id=row.row_id,
+            status="N/A",
+            reason=unsupported_reason,
+            duration_sec=0.0,
+            hint="unsupported_case_precheck",
+            log_path=str(log_path),
+        )
+
+    if row.requires_gpu and not gpu_available:
+        return CaseResult(
+            python_version=python_version,
+            row_id=row.row_id,
+            status="N/A",
+            reason="gpu_unavailable",
+            duration_sec=0.0,
+            hint="nvidia-smi unavailable or failed",
+            log_path=str(log_path),
+        )
+
+    env_dir = repo_root / ".matrix-envs" / safe_slug(f"py{python_version}-{row.row_id}")
+    sync_ok, sync_hint = sync_case_environment(
+        repo_root,
+        python_version,
+        row,
+        env_dir,
+        log_path=log_path,
+    )
+    if not sync_ok:
+        return CaseResult(
+            python_version=python_version,
+            row_id=row.row_id,
+            status="FAIL",
+            reason="dependency_sync_failed",
+            duration_sec=round(time.monotonic() - start, 3),
+            hint=sync_hint,
+            log_path=str(log_path),
+        )
+
+    cmd, process_timeout = build_offline_command(
+        python_version, row, sample_audio, timeout_sec
+    )
+    env = os.environ.copy()
+    env["UV_PROJECT_ENVIRONMENT"] = str(env_dir)
+    if row.requires_gpu:
+        env.pop("CUDA_VISIBLE_DEVICES", None)
+    else:
+        env["CUDA_VISIBLE_DEVICES"] = ""
+    try:
+        proc = run_command(
+            cmd,
+            cwd=repo_root,
+            env=env,
+            timeout=process_timeout,
+            log_path=log_path,
+            log_section="offline",
+        )
+    except subprocess.TimeoutExpired as exc:
+        return CaseResult(
+            python_version=python_version,
+            row_id=row.row_id,
+            status="FAIL",
+            reason="offline_timeout",
+            duration_sec=round(time.monotonic() - start, 3),
+            hint=tail_text((exc.stderr or "") if isinstance(exc.stderr, str) else ""),
+            log_path=str(log_path),
+        )
+
+    hint = tail_text(proc.stderr or proc.stdout)
+    if proc.returncode == 0:
+        return CaseResult(
+            python_version=python_version,
+            row_id=row.row_id,
+            status="PASS",
+            reason="ok",
+            duration_sec=round(time.monotonic() - start, 3),
+            hint=hint,
+            log_path=str(log_path),
+        )
+
+    reason = "offline_timeout" if proc.returncode == 124 else "offline_run_failed"
+    return CaseResult(
+        python_version=python_version,
+        row_id=row.row_id,
+        status="FAIL",
+        reason=reason,
+        duration_sec=round(time.monotonic() - start, 3),
+        hint=hint,
+        log_path=str(log_path),
+    )
+
+
+def print_summary(results: list[CaseResult]) -> None:
+    pass_count = sum(1 for row in results if row.status == "PASS")
+    fail_count = sum(1 for row in results if row.status == "FAIL")
+    na_count = sum(1 for row in results if row.status == "N/A")
+    if CONSOLE is None:
+        print("\n[matrix] results")
+        print("python | row | status | reason | duration_s")
+        print("---|---|---|---|---")
+        for result in results:
+            print(
+                f"{result.python_version} | {result.row_id} | {result.status} | "
+                f"{result.reason} | {result.duration_sec:.3f}"
+            )
+        print(
+            f"\n[matrix] summary pass={pass_count} fail={fail_count} "
+            f"na={na_count} total={len(results)}"
+        )
+    else:
+        table = Table(title="Support Matrix Results")
+        table.add_column("Python", style="cyan", no_wrap=True)
+        table.add_column("Row", style="white")
+        table.add_column("Status", no_wrap=True)
+        table.add_column("Reason")
+        table.add_column("Duration (s)", justify="right", no_wrap=True)
+        for result in results:
+            table.add_row(
+                result.python_version,
+                result.row_id,
+                f"[{status_style(result.status)}]{result.status}[/{status_style(result.status)}]",
+                result.reason,
+                f"{result.duration_sec:.3f}",
+            )
+        CONSOLE.print()
+        CONSOLE.print(table)
+        CONSOLE.print(
+            f"[bold]Summary[/bold] "
+            f"pass=[green]{pass_count}[/green] "
+            f"fail=[bold red]{fail_count}[/bold red] "
+            f"na=[yellow]{na_count}[/yellow] "
+            f"total={len(results)}"
+        )
+
+    diagnostics = [row for row in results if row.status in {"FAIL", "N/A"} and row.hint]
+    if diagnostics:
+        if CONSOLE is None:
+            print("\n[matrix] diagnostics (failed/n-a cases)")
+            for row in diagnostics:
+                print(
+                    f"- py={row.python_version} row={row.row_id} "
+                    f"status={row.status} reason={row.reason}"
+                )
+                print(f"  hint: {row.hint}")
+                if row.log_path:
+                    print(f"  log: {row.log_path}")
+        else:
+            diagnostics_table = Table(title="Diagnostics (FAIL / N/A)")
+            diagnostics_table.add_column("Case", style="cyan")
+            diagnostics_table.add_column("Status", no_wrap=True)
+            diagnostics_table.add_column("Reason")
+            diagnostics_table.add_column("Hint")
+            diagnostics_table.add_column("Log")
+            for row in diagnostics:
+                diagnostics_table.add_row(
+                    f"py={row.python_version} {row.row_id}",
+                    f"[{status_style(row.status)}]{row.status}[/{status_style(row.status)}]",
+                    row.reason,
+                    row.hint,
+                    row.log_path,
+                )
+            CONSOLE.print()
+            CONSOLE.print(diagnostics_table)
+
+
+def main() -> int:
+    args = parse_args()
+    if args.timeout_sec <= 0:
+        print("[matrix] error: --timeout-sec must be > 0", file=sys.stderr)
+        return 1
+
+    repo_root = Path(__file__).resolve().parents[1]
+    logs_dir = (repo_root / args.logs_dir).resolve()
+    logs_dir.mkdir(parents=True, exist_ok=True)
+    print_line(f"[matrix] repo_root={repo_root}", style="cyan")
+    print_line(f"[matrix] timeout_sec={args.timeout_sec}", style="cyan")
+    print_line(f"[matrix] logs_dir={logs_dir}", style="cyan")
+
+    try:
+        sample_audio = download_sample(repo_root)
+    except Exception as exc:  # pragma: no cover - straightforward failure path
+        if CONSOLE is None:
+            print(f"[matrix] sample_download_failed: {exc}", file=sys.stderr)
+        else:
+            CONSOLE.print(
+                f"[matrix] sample_download_failed: {exc}",
+                style="bold red",
+                highlight=False,
+            )
+        return 1
+    print_line(f"[matrix] sample_audio={sample_audio}", style="cyan")
+
+    gpu_available = detect_gpu_available()
+    print_line(f"[matrix] gpu_available={gpu_available}", style="cyan")
+
+    results: list[CaseResult] = []
+    for python_version in PYTHON_VERSIONS:
+        for row in CASES:
+            print_line(
+                f"\n[matrix] running py={python_version} row={row.row_id}", style="blue"
+            )
+            result = run_case(
+                repo_root=repo_root,
+                python_version=python_version,
+                row=row,
+                sample_audio=sample_audio,
+                timeout_sec=args.timeout_sec,
+                gpu_available=gpu_available,
+                logs_dir=logs_dir,
+            )
+            result = apply_expected_failure_policy(result)
+            results.append(result)
+            print_line(
+                f"[matrix] {result.status} py={result.python_version} "
+                f"row={result.row_id} reason={result.reason} duration={result.duration_sec:.3f}s",
+                style=status_style(result.status),
+            )
+            if result.log_path:
+                print_line(f"[matrix] log={result.log_path}", style="dim")
+
+    print_summary(results)
+    fail_count = sum(1 for row in results if row.status == "FAIL")
+    return 1 if fail_count else 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/sync_extension.py

@@ -1,9 +1,10 @@
 """Copy core files from web directory to Chrome extension directory."""
 
-import shutil
 import os
+import shutil
 from pathlib import Path
 
+
 def sync_extension_files():
     
     web_dir = Path("whisperlivekit/web")

test_backend_offline.py

@@ -0,0 +1,803 @@
+#!/usr/bin/env python3
+"""
+Offline test harness and benchmark suite for WhisperLiveKit backends.
+
+Simulates a client-server session by feeding audio files as PCM bytes through
+the full AudioProcessor pipeline (the same path used by the WebSocket server),
+without needing a browser or microphone.
+
+Computes WER (Word Error Rate) and timestamp accuracy when ground truth
+transcript files (.transcript.json) are available alongside audio files.
+
+Usage:
+    # Test with a single audio file:
+    python test_backend_offline.py --backend faster-whisper --audio audio_tests/00_00_07_english_1_speaker.wav
+
+    # Test all files in audio_tests/:
+    python test_backend_offline.py --backend faster-whisper --no-realtime
+
+    # Override streaming policy:
+    python test_backend_offline.py --backend faster-whisper --policy simulstreaming --no-realtime
+
+    # Multi-backend benchmark (auto-detects all installed backends):
+    python test_backend_offline.py --benchmark --no-realtime
+
+    # Export results as JSON:
+    python test_backend_offline.py --benchmark --no-realtime --json results.json
+
+    # Insert silence for testing silence handling:
+    python test_backend_offline.py --backend faster-whisper --insert-silence 3.0 2.0
+"""
+
+import argparse
+import asyncio
+import json
+import logging
+import sys
+import time
+import urllib.request
+from pathlib import Path
+from dataclasses import dataclass, asdict, field
+from typing import List, Optional
+
+import numpy as np
+
+logging.basicConfig(
+    level=logging.WARNING,
+    format="%(asctime)s %(levelname)s %(name)s: %(message)s",
+)
+logger = logging.getLogger("test_offline")
+logger.setLevel(logging.INFO)
+
+SAMPLE_RATE = 16000
+JFK_WAV_URL = "https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav"
+CACHE_DIR = Path(__file__).parent / ".test_cache"
+AUDIO_TESTS_DIR = Path(__file__).parent / "audio_tests"
+AUDIO_EXTENSIONS = {".wav", ".mp3", ".flac", ".ogg", ".m4a"}
+
+
+@dataclass
+class WordTimestamp:
+    """Word with its start/end time."""
+    word: str
+    start: float
+    end: float
+
+
+@dataclass
+class TestResult:
+    """Structured result from a single test run."""
+    audio_file: str
+    audio_duration_s: float
+    backend: str
+    policy: str
+    language: str
+    chunk_ms: int
+    realtime_pacing: bool
+    # Timing
+    processing_time_s: float
+    rtf: float  # real-time factor
+    # Transcription output
+    transcription: str
+    n_lines: int
+    n_responses: int
+    # WER metrics (None if no ground truth)
+    wer: Optional[float] = None
+    wer_details: Optional[dict] = None
+    # Timestamp accuracy (None if no ground truth)
+    timestamp_mae: Optional[float] = None
+    timestamp_max_delta: Optional[float] = None
+    timestamp_median_delta: Optional[float] = None
+    # Word-level timestamps
+    word_timestamps: List[WordTimestamp] = field(default_factory=list)
+    # Raw last response
+    last_response: Optional[dict] = None
+
+
+def download_sample_audio() -> Path:
+    """Download the jfk.wav sample if not cached."""
+    CACHE_DIR.mkdir(exist_ok=True)
+    path = CACHE_DIR / "jfk.wav"
+    if not path.exists():
+        logger.info(f"Downloading sample audio to {path} ...")
+        urllib.request.urlretrieve(JFK_WAV_URL, path)
+        logger.info("Done.")
+    return path
+
+
+def load_audio(path: str) -> np.ndarray:
+    """Load audio file as float32 mono 16kHz numpy array.
+
+    Supports WAV, FLAC (via soundfile) and MP3, OGG, M4A (via librosa).
+    """
+    ext = Path(path).suffix.lower()
+    if ext in (".mp3", ".ogg", ".m4a"):
+        import librosa
+        audio, _ = librosa.load(path, sr=SAMPLE_RATE, mono=True)
+        return audio.astype(np.float32)
+
+    import soundfile as sf
+    audio, sr = sf.read(path, dtype="float32")
+    if audio.ndim > 1:
+        audio = audio.mean(axis=1)
+    if sr != SAMPLE_RATE:
+        import librosa
+        audio = librosa.resample(audio, orig_sr=sr, target_sr=SAMPLE_RATE)
+    return audio
+
+
+def insert_silence(audio: np.ndarray, silence_sec: float, position_sec: float) -> np.ndarray:
+    """Insert silence into audio at a given position.
+
+    Args:
+        audio: Float32 mono audio array at SAMPLE_RATE.
+        silence_sec: Duration of silence to insert in seconds.
+        position_sec: Position in seconds where silence starts.
+    Returns:
+        New audio array with silence inserted.
+    """
+    pos_samples = int(position_sec * SAMPLE_RATE)
+    silence_samples = int(silence_sec * SAMPLE_RATE)
+    pos_samples = min(pos_samples, len(audio))
+    silence = np.zeros(silence_samples, dtype=np.float32)
+    return np.concatenate([audio[:pos_samples], silence, audio[pos_samples:]])
+
+
+def float32_to_s16le_bytes(audio: np.ndarray) -> bytes:
+    """Convert float32 audio to s16le PCM bytes (what the browser sends)."""
+    return (audio * 32768).clip(-32768, 32767).astype(np.int16).tobytes()
+
+
+def create_engine(
+    backend: str, model_size: str, lan: str,
+    diarization: bool = False,
+    diarization_backend: str = "",
+    vac: bool = True,
+    policy: str = "",
+):
+    """Create a TranscriptionEngine with the given backend config."""
+    import gc
+    from whisperlivekit.core import TranscriptionEngine
+
+    # Reset singleton so we get a fresh instance
+    TranscriptionEngine._instance = None
+    TranscriptionEngine._initialized = False
+    gc.collect()
+
+    kwargs = dict(
+        backend=backend,
+        lan=lan,
+        pcm_input=True,
+        vac=vac,
+        transcription=True,
+        diarization=diarization,
+    )
+    if diarization_backend:
+        kwargs["diarization_backend"] = diarization_backend
+    if model_size:
+        kwargs["model_size"] = model_size
+    if policy:
+        kwargs["backend_policy"] = policy
+
+    return TranscriptionEngine(**kwargs)
+
+
+def _extract_text_from_response(response_dict: dict) -> str:
+    """Extract full transcription text from a FrontData dict."""
+    def _strip_or_empty(value: object) -> str:
+        return value.strip() if isinstance(value, str) else ""
+
+    segments = response_dict.get("lines", [])
+    full_text = " ".join(
+        text
+        for seg in segments
+        if isinstance(seg, dict)
+        for text in [_strip_or_empty(seg.get("text"))]
+        if text
+    )
+    buf = _strip_or_empty(response_dict.get("buffer_transcription"))
+    if buf:
+        full_text = f"{full_text} {buf}".strip() if full_text else buf
+    return full_text
+
+
+async def run_test(
+    engine, audio: np.ndarray, chunk_ms: int, realtime: bool,
+    audio_file: str = "", backend: str = "", policy: str = "", lan: str = "",
+) -> TestResult:
+    """
+    Simulate a client session through the full AudioProcessor pipeline.
+
+    1. Create AudioProcessor (one per "client session")
+    2. Start async pipeline (transcription_processor, results_formatter, etc.)
+    3. Feed audio as PCM bytes in timed chunks
+    4. Collect and display FrontData responses
+    5. Signal EOF and cleanup
+    """
+    from whisperlivekit.audio_processor import AudioProcessor
+
+    chunk_samples = int(SAMPLE_RATE * chunk_ms / 1000)
+    total_samples = len(audio)
+    audio_duration = total_samples / SAMPLE_RATE
+
+    logger.info(
+        f"Audio: {audio_duration:.2f}s | "
+        f"Chunk: {chunk_ms}ms ({chunk_samples} samples) | "
+        f"Steps: {total_samples // chunk_samples + 1} | "
+        f"Realtime: {realtime}"
+    )
+
+    # --- Server side: create processor and start pipeline ---
+    processor = AudioProcessor(transcription_engine=engine)
+    results_generator = await processor.create_tasks()
+
+    # Collect results in background (like handle_websocket_results)
+    all_responses = []
+    response_count = 0
+    last_printed_text = ""
+
+    async def collect_results():
+        nonlocal response_count, last_printed_text
+        async for response in results_generator:
+            all_responses.append(response)
+            response_count += 1
+            d = response.to_dict()
+
+            # Only print when transcription text actually changes
+            current_text = _extract_text_from_response(d)
+            if current_text and current_text != last_printed_text:
+                buf = d.get("buffer_transcription")
+                buf = buf.strip() if isinstance(buf, str) else ""
+                committed = current_text
+                if buf and committed.endswith(buf):
+                    committed = committed[:-len(buf)].strip()
+
+                # Show committed text + buffer separately
+                display = committed
+                if buf:
+                    display = f"{committed} \033[90m{buf}\033[0m" if committed else f"\033[90m{buf}\033[0m"
+                print(f"  > {display}", flush=True)
+                last_printed_text = current_text
+
+    result_task = asyncio.create_task(collect_results())
+
+    # --- Client side: feed audio as PCM bytes ---
+    t_start = time.time()
+
+    for offset in range(0, total_samples, chunk_samples):
+        chunk = audio[offset : offset + chunk_samples]
+        pcm_bytes = float32_to_s16le_bytes(chunk)
+        await processor.process_audio(pcm_bytes)
+        if realtime:
+            await asyncio.sleep(chunk_ms / 1000)
+
+    feed_elapsed = time.time() - t_start
+
+    logger.info(f"Audio fed in {feed_elapsed:.2f}s. Signaling EOF...")
+
+    # Signal end of audio (like client disconnect / empty message)
+    await processor.process_audio(None)
+
+    # Wait for pipeline to drain completely
+    try:
+        await asyncio.wait_for(result_task, timeout=120.0)
+    except asyncio.TimeoutError:
+        logger.warning("Timed out waiting for results. Proceeding with cleanup.")
+        result_task.cancel()
+        try:
+            await result_task
+        except asyncio.CancelledError:
+            pass
+
+    # --- Capture word-level timestamps before cleanup ---
+    word_timestamps = []
+    try:
+        state = await processor.get_current_state()
+        for token in state.tokens:
+            if hasattr(token, 'start') and hasattr(token, 'text') and token.text:
+                word_timestamps.append(WordTimestamp(
+                    word=token.text.strip(),
+                    start=round(token.start, 3),
+                    end=round(token.end, 3),
+                ))
+    except Exception as e:
+        logger.warning(f"Could not capture word timestamps: {e}")
+
+    # Cleanup
+    await processor.cleanup()
+
+    total_elapsed = time.time() - t_start
+
+    # --- Build result ---
+    transcription = ""
+    n_lines = 0
+    last_response_dict = None
+
+    if all_responses:
+        last = all_responses[-1].to_dict()
+        last_response_dict = last
+        n_lines = len(last.get("lines", []))
+        transcription = _extract_text_from_response(last)
+
+    # --- Compute WER and timestamp accuracy against ground truth ---
+    from whisperlivekit.metrics import compute_wer, compute_timestamp_accuracy
+
+    wer_val = None
+    wer_details = None
+    ts_mae = None
+    ts_max_delta = None
+    ts_median_delta = None
+
+    gt_path = Path(audio_file).with_suffix(".transcript.json")
+    if not gt_path.exists():
+        gt_path = AUDIO_TESTS_DIR / gt_path
+    gt = None
+    if gt_path.exists():
+        with open(gt_path) as f:
+            gt = json.load(f)
+
+        # WER
+        gt_text = " ".join(w["word"] for w in gt)
+        wer_result = compute_wer(gt_text, transcription)
+        wer_val = round(wer_result["wer"], 4)
+        wer_details = wer_result
+
+        # Timestamp accuracy
+        if word_timestamps:
+            pred_dicts = [{"word": wt.word, "start": wt.start, "end": wt.end} for wt in word_timestamps]
+            ts_result = compute_timestamp_accuracy(pred_dicts, gt)
+            ts_mae = ts_result["mae_start"]
+            ts_max_delta = ts_result["max_delta_start"]
+            ts_median_delta = ts_result["median_delta_start"]
+
+    result = TestResult(
+        audio_file=audio_file,
+        audio_duration_s=round(audio_duration, 2),
+        backend=backend,
+        policy=policy,
+        language=lan,
+        chunk_ms=chunk_ms,
+        realtime_pacing=realtime,
+        processing_time_s=round(total_elapsed, 2),
+        rtf=round(total_elapsed / audio_duration, 2),
+        transcription=transcription,
+        n_lines=n_lines,
+        n_responses=response_count,
+        wer=wer_val,
+        wer_details=wer_details,
+        timestamp_mae=round(ts_mae, 3) if ts_mae is not None else None,
+        timestamp_max_delta=round(ts_max_delta, 3) if ts_max_delta is not None else None,
+        timestamp_median_delta=round(ts_median_delta, 3) if ts_median_delta is not None else None,
+        word_timestamps=word_timestamps,
+        last_response=last_response_dict,
+    )
+
+    # --- Print summary ---
+    print(f"\n{'=' * 60}")
+    print(f"RESULT: {audio_file}")
+    print(f"{'=' * 60}")
+    print(f"Transcription: {transcription}")
+    print(f"Lines: {n_lines} | Responses: {response_count}")
+    print(f"Audio: {audio_duration:.2f}s | Time: {total_elapsed:.2f}s | RTF: {result.rtf:.2f}x")
+
+    if wer_val is not None:
+        print(f"WER: {wer_val:.2%} (S={wer_details['substitutions']} I={wer_details['insertions']} D={wer_details['deletions']})")
+
+    # Print word timestamps if available
+    if word_timestamps:
+        print(f"\nWord timestamps ({len(word_timestamps)} words):")
+        for wt in word_timestamps:
+            print(f"  [{wt.start:6.2f} - {wt.end:6.2f}] {wt.word}")
+
+        # Detailed comparison with ground truth
+        if gt:
+            print(f"\n  vs Ground truth ({len(gt)} words):")
+            max_words = max(len(word_timestamps), len(gt))
+            for i in range(max_words):
+                pred = word_timestamps[i] if i < len(word_timestamps) else None
+                ref = gt[i] if i < len(gt) else None
+                p_str = f"[{pred.start:5.2f}-{pred.end:5.2f}] {pred.word:<15}" if pred else " " * 30
+                r_str = f"[{ref['start']:5.2f}-{ref['end']:5.2f}] {ref['word']:<15}" if ref else ""
+                delta = ""
+                if pred and ref:
+                    d = pred.start - ref['start']
+                    delta = f"  Δstart={d:+.2f}"
+                print(f"  {p_str}  |  {r_str}{delta}")
+
+        if ts_mae is not None:
+            print(f"\n  Timestamp stats: MAE={ts_mae:.3f}s  max|Δ|={ts_max_delta:.3f}s  median|Δ|={ts_median_delta:.3f}s")
+
+    print(f"{'=' * 60}")
+
+    return result
+
+
+def discover_audio_files(directory: str) -> List[Path]:
+    """Find all supported audio files in directory."""
+    d = Path(directory)
+    files = sorted(
+        p for p in d.iterdir()
+        if p.is_file() and p.suffix.lower() in AUDIO_EXTENSIONS
+    )
+    return files
+
+
+async def run_all_tests(
+    engine, audio_files: List[Path], chunk_ms: int, realtime: bool,
+    backend: str, policy: str, lan: str, max_duration: float = 60.0,
+    silence_insertions: Optional[List[List[float]]] = None,
+) -> List[TestResult]:
+    """Run tests on multiple audio files sequentially."""
+    results = []
+    for audio_path in audio_files:
+        # Detect language from filename if "french" in name
+        file_lan = lan
+        if "french" in audio_path.name.lower() and lan == "en":
+            file_lan = "fr"
+            logger.info(f"Auto-detected language 'fr' from filename")
+
+        audio = load_audio(str(audio_path))
+
+        # Insert silence segments (applied in reverse position order to keep offsets valid)
+        if silence_insertions:
+            for secs, at_sec in sorted(silence_insertions, key=lambda x: x[1], reverse=True):
+                logger.info(f"Inserting {secs:.1f}s silence at {at_sec:.1f}s")
+                audio = insert_silence(audio, secs, at_sec)
+
+        duration = len(audio) / SAMPLE_RATE
+
+        if duration > max_duration:
+            logger.info(f"Skipping {audio_path.name} ({duration:.0f}s > {max_duration:.0f}s max)")
+            continue
+
+        print(f"\n{'#' * 60}")
+        print(f"# Testing: {audio_path.name} ({duration:.1f}s)")
+        print(f"{'#' * 60}")
+
+        result = await run_test(
+            engine, audio, chunk_ms, realtime,
+            audio_file=audio_path.name, backend=backend, policy=policy, lan=file_lan,
+        )
+        results.append(result)
+
+    return results
+
+
+def print_benchmark_summary(results: List[TestResult]):
+    """Print a tabular summary of all test results."""
+    print(f"\n{'=' * 110}")
+    print("BENCHMARK SUMMARY")
+    print(f"{'=' * 110}")
+    print(
+        f"{'File':<40} {'Duration':>8} {'Time':>8} {'RTF':>6} "
+        f"{'WER':>7} {'MAE(s)':>7} {'Lines':>5}"
+    )
+    print(f"{'-' * 110}")
+    for r in results:
+        wer_str = f"{r.wer:.2%}" if r.wer is not None else "  -"
+        mae_str = f"{r.timestamp_mae:.3f}" if r.timestamp_mae is not None else "  -"
+        print(
+            f"{r.audio_file:<40} {r.audio_duration_s:>7.1f}s {r.processing_time_s:>7.1f}s "
+            f"{r.rtf:>5.2f}x {wer_str:>7} {mae_str:>7} {r.n_lines:>5}"
+        )
+    print(f"{'-' * 110}")
+    total_audio = sum(r.audio_duration_s for r in results)
+    total_time = sum(r.processing_time_s for r in results)
+    avg_rtf = total_time / total_audio if total_audio > 0 else 0
+    wer_vals = [r.wer for r in results if r.wer is not None]
+    avg_wer_str = f"{sum(wer_vals)/len(wer_vals):.2%}" if wer_vals else "  -"
+    mae_vals = [r.timestamp_mae for r in results if r.timestamp_mae is not None]
+    avg_mae_str = f"{sum(mae_vals)/len(mae_vals):.3f}" if mae_vals else "  -"
+    print(
+        f"{'TOTAL/AVG':<40} {total_audio:>7.1f}s {total_time:>7.1f}s "
+        f"{avg_rtf:>5.2f}x {avg_wer_str:>7} {avg_mae_str:>7}"
+    )
+    print(f"{'=' * 110}")
+
+    # Print transcription excerpts
+    print(f"\nTRANSCRIPTIONS:")
+    print(f"{'-' * 110}")
+    for r in results:
+        excerpt = r.transcription[:120] + "..." if len(r.transcription) > 120 else r.transcription
+        print(f"  {r.audio_file}:")
+        print(f"    {excerpt}")
+    print(f"{'=' * 110}")
+
+
+def detect_available_backends() -> List[dict]:
+    """Probe which backends can be imported and return (backend, policy) combos.
+
+    Returns list of dicts with keys: backend, policy, description.
+    """
+    combos = []
+
+    # faster-whisper
+    try:
+        import faster_whisper  # noqa: F401
+        combos.append({"backend": "faster-whisper", "policy": "localagreement", "description": "faster-whisper + LocalAgreement"})
+        combos.append({"backend": "faster-whisper", "policy": "simulstreaming", "description": "faster-whisper + SimulStreaming"})
+    except ImportError:
+        pass
+
+    # mlx-whisper (macOS only)
+    try:
+        import mlx_whisper  # noqa: F401
+        combos.append({"backend": "mlx-whisper", "policy": "localagreement", "description": "mlx-whisper + LocalAgreement"})
+        combos.append({"backend": "mlx-whisper", "policy": "simulstreaming", "description": "mlx-whisper + SimulStreaming"})
+    except ImportError:
+        pass
+
+    # openai-whisper
+    try:
+        import whisper  # noqa: F401
+        combos.append({"backend": "whisper", "policy": "localagreement", "description": "openai-whisper + LocalAgreement"})
+        combos.append({"backend": "whisper", "policy": "simulstreaming", "description": "openai-whisper + SimulStreaming"})
+    except ImportError:
+        pass
+
+    # voxtral-mlx
+    try:
+        from whisperlivekit.voxtral_mlx import VoxtralMLXModel  # noqa: F401
+        combos.append({"backend": "voxtral-mlx", "policy": "voxtral", "description": "voxtral-mlx (MLX)"})
+    except ImportError:
+        pass
+
+    # voxtral (HuggingFace)
+    try:
+        from transformers import AutoModelForSpeechSeq2Seq  # noqa: F401
+        combos.append({"backend": "voxtral", "policy": "voxtral", "description": "voxtral (HuggingFace)"})
+    except ImportError:
+        pass
+
+    return combos
+
+
+def print_cross_backend_comparison(all_results: List[TestResult]):
+    """Print a comparison table across backends and policies."""
+    print(f"\n{'=' * 110}")
+    print("CROSS-BACKEND BENCHMARK COMPARISON")
+    print(f"{'=' * 110}")
+    print(
+        f"{'Backend':<18} {'Policy':<16} {'File':<30} "
+        f"{'WER':>7} {'RTF':>6} {'MAE(s)':>7} {'MaxΔ(s)':>8}"
+    )
+    print(f"{'-' * 110}")
+
+    for r in all_results:
+        wer_str = f"{r.wer:.2%}" if r.wer is not None else "  -"
+        rtf_str = f"{r.rtf:.2f}x"
+        mae_str = f"{r.timestamp_mae:.3f}" if r.timestamp_mae is not None else "  -"
+        max_str = f"{r.timestamp_max_delta:.3f}" if r.timestamp_max_delta is not None else "  -"
+        # Truncate filename for readability
+        fname = r.audio_file[:28] + ".." if len(r.audio_file) > 30 else r.audio_file
+        print(
+            f"{r.backend:<18} {r.policy:<16} {fname:<30} "
+            f"{wer_str:>7} {rtf_str:>6} {mae_str:>7} {max_str:>8}"
+        )
+
+    print(f"{'-' * 110}")
+
+    # Per-backend averages
+    from collections import defaultdict
+    by_combo = defaultdict(list)
+    for r in all_results:
+        by_combo[(r.backend, r.policy)].append(r)
+
+    print(f"\n{'Backend':<18} {'Policy':<16} {'Avg WER':>8} {'Avg RTF':>8} {'Avg MAE':>8} {'Files':>6}")
+    print(f"{'-' * 80}")
+    for (backend, policy), group in sorted(by_combo.items()):
+        wer_vals = [r.wer for r in group if r.wer is not None]
+        rtf_vals = [r.rtf for r in group]
+        mae_vals = [r.timestamp_mae for r in group if r.timestamp_mae is not None]
+        avg_wer = f"{sum(wer_vals)/len(wer_vals):.2%}" if wer_vals else "  -"
+        avg_rtf = f"{sum(rtf_vals)/len(rtf_vals):.2f}x"
+        avg_mae = f"{sum(mae_vals)/len(mae_vals):.3f}" if mae_vals else "  -"
+        print(
+            f"{backend:<18} {policy:<16} {avg_wer:>8} {avg_rtf:>8} {avg_mae:>8} {len(group):>6}"
+        )
+    print(f"{'=' * 110}")
+
+
+def _quiet_loggers(verbose: bool):
+    """Set internal module log levels to reduce noise."""
+    if verbose:
+        logging.getLogger().setLevel(logging.DEBUG)
+    else:
+        for mod in (
+            "whisperlivekit.audio_processor", "whisperlivekit.simul_whisper",
+            "whisperlivekit.tokens_alignment", "whisperlivekit.simul_whisper.align_att_base",
+            "whisperlivekit.simul_whisper.simul_whisper",
+        ):
+            logging.getLogger(mod).setLevel(logging.WARNING)
+
+
+async def run_benchmark(
+    audio_files: List[Path], chunk_ms: int, realtime: bool,
+    model_size: str, lan: str, max_duration: float, vac: bool,
+    verbose: bool,
+) -> List[TestResult]:
+    """Run benchmark across all available backend+policy combinations."""
+    combos = detect_available_backends()
+    if not combos:
+        logger.error("No backends available. Install at least one ASR backend.")
+        return []
+
+    logger.info(f"Detected {len(combos)} backend+policy combinations:")
+    for c in combos:
+        logger.info(f"  - {c['description']}")
+
+    all_results = []
+    for i, combo in enumerate(combos, 1):
+        backend = combo["backend"]
+        policy = combo["policy"]
+        desc = combo["description"]
+
+        print(f"\n{'*' * 70}")
+        print(f"* BENCHMARK {i}/{len(combos)}: {desc}")
+        print(f"{'*' * 70}")
+
+        try:
+            engine = create_engine(
+                backend, model_size, lan, vac=vac, policy=policy,
+            )
+            _quiet_loggers(verbose)
+
+            results = await run_all_tests(
+                engine, audio_files, chunk_ms, realtime,
+                backend=backend, policy=policy, lan=lan,
+                max_duration=max_duration,
+            )
+            all_results.extend(results)
+        except Exception as e:
+            logger.error(f"Failed to run {desc}: {e}")
+            import traceback
+            traceback.print_exc()
+
+    return all_results
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Offline backend test harness (AudioProcessor-level)"
+    )
+    parser.add_argument(
+        "--backend", default="faster-whisper",
+        help="Backend: voxtral, voxtral-mlx, auto, faster-whisper, mlx-whisper, whisper.",
+    )
+    parser.add_argument(
+        "--policy", default="",
+        help="Override backend policy: localagreement, simulstreaming, voxtral.",
+    )
+    parser.add_argument(
+        "--audio", default=None,
+        help="Path to a single audio file (WAV, MP3, FLAC, etc.).",
+    )
+    parser.add_argument(
+        "--audio-dir", default=None,
+        help="Directory of audio files to test. Defaults to audio_tests/ if neither --audio nor --audio-dir given.",
+    )
+    parser.add_argument(
+        "--chunk-ms", type=int, default=100,
+        help="Chunk size in milliseconds (simulates real-time interval).",
+    )
+    parser.add_argument(
+        "--model", default="", dest="model_size",
+        help="Model size or HF repo ID.",
+    )
+    parser.add_argument("--lan", default="en", help="Language code.")
+    parser.add_argument(
+        "--no-realtime", action="store_true",
+        help="Skip real-time pacing between chunks (faster but less realistic).",
+    )
+    parser.add_argument(
+        "--no-vac", action="store_true",
+        help="Disable Voice Activity Classification (send all audio without silence filtering).",
+    )
+    parser.add_argument(
+        "--diarization", action="store_true",
+        help="Enable speaker diarization.",
+    )
+    parser.add_argument(
+        "--diarization-backend",
+        default="",
+        choices=["diart", "sortformer"],
+        help="Diarization backend when --diarization is enabled.",
+    )
+    parser.add_argument(
+        "--benchmark", action="store_true",
+        help="Run benchmark across all detected backend+policy combinations.",
+    )
+    parser.add_argument(
+        "--json", default=None, dest="json_output",
+        help="Write structured JSON results to this file.",
+    )
+    parser.add_argument(
+        "--max-duration", type=float, default=60.0,
+        help="Skip audio files longer than this many seconds (default: 60).",
+    )
+    parser.add_argument(
+        "--insert-silence", nargs=2, type=float, metavar=("SECS", "AT_SEC"),
+        action="append", default=[],
+        help="Insert SECS of silence at AT_SEC position. Can be repeated. "
+             "E.g.: --insert-silence 3.0 2.0 --insert-silence 5.0 7.0",
+    )
+    parser.add_argument(
+        "-v", "--verbose", action="store_true",
+        help="Show debug-level logs from all components.",
+    )
+    args = parser.parse_args()
+
+    realtime = not args.no_realtime
+    vac = not args.no_vac
+
+    # Resolve audio file(s)
+    if args.audio:
+        audio_files = [Path(args.audio)]
+    elif args.audio_dir:
+        audio_files = discover_audio_files(args.audio_dir)
+    elif AUDIO_TESTS_DIR.is_dir():
+        audio_files = discover_audio_files(str(AUDIO_TESTS_DIR))
+    else:
+        # Fall back to jfk.wav download
+        audio_files = [download_sample_audio()]
+
+    if not audio_files:
+        logger.error("No audio files found.")
+        sys.exit(1)
+
+    logger.info(f"Audio files: {[f.name for f in audio_files]}")
+
+    if args.benchmark:
+        # --- Multi-backend benchmark mode ---
+        all_results = asyncio.run(
+            run_benchmark(
+                audio_files, args.chunk_ms, realtime,
+                args.model_size, args.lan, args.max_duration, vac,
+                args.verbose,
+            )
+        )
+        if all_results:
+            print_cross_backend_comparison(all_results)
+        results = all_results
+    else:
+        # --- Single-backend mode ---
+        policy = args.policy
+        logger.info(f"Creating {args.backend} engine...")
+        engine = create_engine(
+            args.backend, args.model_size, args.lan,
+            diarization=args.diarization,
+            diarization_backend=args.diarization_backend,
+            vac=vac,
+            policy=policy,
+        )
+        logger.info("Engine ready.")
+
+        _quiet_loggers(args.verbose)
+
+        results = asyncio.run(
+            run_all_tests(
+                engine, audio_files, args.chunk_ms, realtime,
+                args.backend, policy, args.lan,
+                max_duration=args.max_duration,
+                silence_insertions=args.insert_silence or None,
+            )
+        )
+
+        if len(results) > 1:
+            print_benchmark_summary(results)
+
+    # JSON output
+    if args.json_output and results:
+        json_results = []
+        for r in results:
+            d = asdict(r)
+            d.pop("last_response", None)  # too verbose for summary
+            json_results.append(d)
+        Path(args.json_output).write_text(
+            json.dumps(json_results, indent=2, ensure_ascii=False)
+        )
+        logger.info(f"Results written to {args.json_output}")
+
+
+if __name__ == "__main__":
+    main()

tests/conftest.py

@@ -0,0 +1,58 @@
+"""Shared pytest fixtures for WhisperLiveKit tests."""
+
+import json
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+from whisperlivekit.timed_objects import ASRToken, Silence, Transcript
+
+
+AUDIO_TESTS_DIR = Path(__file__).parent.parent / "audio_tests"
+
+
+@pytest.fixture
+def sample_tokens():
+    """A short sequence of ASRToken objects."""
+    return [
+        ASRToken(start=0.0, end=0.5, text="Hello"),
+        ASRToken(start=0.5, end=1.0, text=" world"),
+        ASRToken(start=1.0, end=1.5, text=" test."),
+    ]
+
+
+@pytest.fixture
+def sample_silence():
+    """A completed silence event."""
+    s = Silence(start=1.5, end=3.0, is_starting=False, has_ended=True)
+    s.compute_duration()
+    return s
+
+
+@pytest.fixture
+def mock_args():
+    """Minimal args namespace for AudioProcessor tests."""
+    return SimpleNamespace(
+        diarization=False,
+        transcription=True,
+        target_language="",
+        vac=False,
+        vac_chunk_size=0.04,
+        min_chunk_size=0.1,
+        pcm_input=True,
+        punctuation_split=False,
+        backend="faster-whisper",
+        backend_policy="localagreement",
+        vad=True,
+    )
+
+
+@pytest.fixture
+def ground_truth_en():
+    """Ground truth transcript for the 7s English audio (if available)."""
+    path = AUDIO_TESTS_DIR / "00_00_07_english_1_speaker.transcript.json"
+    if path.exists():
+        with open(path) as f:
+            return json.load(f)
+    return None

tests/test_audio_processor.py

@@ -0,0 +1,209 @@
+"""Tests for AudioProcessor pipeline with mocked ASR backends.
+
+These tests verify the async audio processing pipeline works correctly
+without requiring any real ASR models to be loaded.
+"""
+
+import asyncio
+from types import SimpleNamespace
+from unittest.mock import patch
+
+import numpy as np
+import pytest
+
+from whisperlivekit.timed_objects import ASRToken, Transcript
+
+
+# ---------------------------------------------------------------------------
+# Mock ASR components
+# ---------------------------------------------------------------------------
+
+class MockASR:
+    """Mock ASR model holder."""
+    sep = " "
+    SAMPLING_RATE = 16000
+
+    def __init__(self):
+        self.transcribe_kargs = {}
+        self.original_language = "en"
+        self.backend_choice = "mock"
+
+    def transcribe(self, audio):
+        return None
+
+
+class MockOnlineProcessor:
+    """Mock online processor that returns canned tokens."""
+    SAMPLING_RATE = 16000
+
+    def __init__(self, asr=None):
+        self.asr = asr or MockASR()
+        self.audio_buffer = np.array([], dtype=np.float32)
+        self.end = 0.0
+        self._call_count = 0
+        self._finished = False
+
+    def insert_audio_chunk(self, audio, audio_stream_end_time):
+        self.audio_buffer = np.append(self.audio_buffer, audio)
+        self.end = audio_stream_end_time
+
+    def process_iter(self, is_last=False):
+        self._call_count += 1
+        # Emit a token on every call when we have audio
+        if len(self.audio_buffer) > 0:
+            t = self._call_count * 0.5
+            return [ASRToken(start=t, end=t + 0.5, text=f"word{self._call_count}")], self.end
+        return [], self.end
+
+    def get_buffer(self):
+        return Transcript(start=None, end=None, text="")
+
+    def start_silence(self):
+        return [], self.end
+
+    def end_silence(self, silence_duration, offset):
+        pass
+
+    def new_speaker(self, change_speaker):
+        pass
+
+    def finish(self):
+        self._finished = True
+        return [], self.end
+
+    def warmup(self, audio, init_prompt=""):
+        pass
+
+
+def _make_pcm_bytes(duration_s=0.1, sample_rate=16000):
+    """Generate silent PCM s16le bytes."""
+    n_samples = int(duration_s * sample_rate)
+    audio = np.zeros(n_samples, dtype=np.float32)
+    return (audio * 32768).clip(-32768, 32767).astype(np.int16).tobytes()
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+@pytest.fixture
+def mock_engine():
+    """Create a mock TranscriptionEngine-like object."""
+    engine = SimpleNamespace(
+        asr=MockASR(),
+        diarization_model=None,
+        translation_model=None,
+        args=SimpleNamespace(
+            diarization=False,
+            transcription=True,
+            target_language="",
+            vac=False,
+            vac_chunk_size=0.04,
+            min_chunk_size=0.1,
+            pcm_input=True,
+            punctuation_split=False,
+            backend="mock",
+            backend_policy="localagreement",
+            vad=True,
+            model_size="base",
+            lan="en",
+        ),
+    )
+    return engine
+
+
+# ---------------------------------------------------------------------------
+# Tests
+# ---------------------------------------------------------------------------
+
+class TestPCMConversion:
+    """Test PCM byte conversion without needing the full pipeline."""
+
+    def test_s16le_roundtrip(self):
+        """Convert float32 → s16le → float32 and verify approximate roundtrip."""
+        original = np.array([0.0, 0.5, -0.5, 1.0, -1.0], dtype=np.float32)
+        s16 = (original * 32768).clip(-32768, 32767).astype(np.int16)
+        pcm_bytes = s16.tobytes()
+        # Direct numpy conversion (same logic as AudioProcessor.convert_pcm_to_float)
+        recovered = np.frombuffer(pcm_bytes, dtype=np.int16).astype(np.float32) / 32768.0
+
+        np.testing.assert_allclose(recovered, original, atol=1 / 32768)
+
+
+@pytest.mark.asyncio
+class TestPipelineBasics:
+    async def test_feed_audio_and_get_responses(self, mock_engine):
+        """Feed audio through the pipeline and verify we get responses."""
+        from whisperlivekit.audio_processor import AudioProcessor
+
+        with patch("whisperlivekit.audio_processor.online_factory", return_value=MockOnlineProcessor()):
+            processor = AudioProcessor(transcription_engine=mock_engine)
+            results_gen = await processor.create_tasks()
+
+            responses = []
+
+            async def collect():
+                async for resp in results_gen:
+                    responses.append(resp)
+
+            task = asyncio.create_task(collect())
+
+            # Feed 2 seconds of audio in 100ms chunks
+            for _ in range(20):
+                await processor.process_audio(_make_pcm_bytes(0.1))
+
+            # Signal EOF
+            await processor.process_audio(None)
+
+            await asyncio.wait_for(task, timeout=10.0)
+            await processor.cleanup()
+
+            # We should have gotten at least one response
+            assert len(responses) > 0
+
+    async def test_eof_terminates_pipeline(self, mock_engine):
+        """Sending None (EOF) should cleanly terminate the pipeline."""
+        from whisperlivekit.audio_processor import AudioProcessor
+
+        with patch("whisperlivekit.audio_processor.online_factory", return_value=MockOnlineProcessor()):
+            processor = AudioProcessor(transcription_engine=mock_engine)
+            results_gen = await processor.create_tasks()
+
+            responses = []
+
+            async def collect():
+                async for resp in results_gen:
+                    responses.append(resp)
+
+            task = asyncio.create_task(collect())
+
+            # Send a small amount of audio then EOF
+            await processor.process_audio(_make_pcm_bytes(0.5))
+            await processor.process_audio(None)
+
+            await asyncio.wait_for(task, timeout=10.0)
+            await processor.cleanup()
+
+            # Pipeline should have terminated without error
+            assert task.done()
+
+    async def test_empty_audio_no_crash(self, mock_engine):
+        """Sending EOF immediately (no audio) should not crash."""
+        from whisperlivekit.audio_processor import AudioProcessor
+
+        with patch("whisperlivekit.audio_processor.online_factory", return_value=MockOnlineProcessor()):
+            processor = AudioProcessor(transcription_engine=mock_engine)
+            results_gen = await processor.create_tasks()
+
+            responses = []
+
+            async def collect():
+                async for resp in results_gen:
+                    responses.append(resp)
+
+            task = asyncio.create_task(collect())
+            await processor.process_audio(None)
+
+            await asyncio.wait_for(task, timeout=10.0)
+            await processor.cleanup()
+            assert task.done()

tests/test_config.py

@@ -0,0 +1,99 @@
+"""Tests for WhisperLiveKitConfig."""
+
+import logging
+from types import SimpleNamespace
+
+import pytest
+
+from whisperlivekit.config import WhisperLiveKitConfig
+
+
+class TestDefaults:
+    def test_default_backend(self):
+        c = WhisperLiveKitConfig()
+        assert c.backend == "auto"
+
+    def test_default_policy(self):
+        c = WhisperLiveKitConfig()
+        assert c.backend_policy == "simulstreaming"
+
+    def test_default_language(self):
+        c = WhisperLiveKitConfig()
+        assert c.lan == "auto"
+
+    def test_default_vac(self):
+        c = WhisperLiveKitConfig()
+        assert c.vac is True
+
+    def test_default_model_size(self):
+        c = WhisperLiveKitConfig()
+        assert c.model_size == "base"
+
+    def test_default_transcription(self):
+        c = WhisperLiveKitConfig()
+        assert c.transcription is True
+        assert c.diarization is False
+
+
+class TestPostInit:
+    def test_en_model_forces_english(self):
+        c = WhisperLiveKitConfig(model_size="tiny.en")
+        assert c.lan == "en"
+
+    def test_en_suffix_with_auto_language(self):
+        c = WhisperLiveKitConfig(model_size="base.en", lan="auto")
+        assert c.lan == "en"
+
+    def test_non_en_model_keeps_language(self):
+        c = WhisperLiveKitConfig(model_size="base", lan="fr")
+        assert c.lan == "fr"
+
+    def test_policy_alias_1(self):
+        c = WhisperLiveKitConfig(backend_policy="1")
+        assert c.backend_policy == "simulstreaming"
+
+    def test_policy_alias_2(self):
+        c = WhisperLiveKitConfig(backend_policy="2")
+        assert c.backend_policy == "localagreement"
+
+    def test_policy_no_alias(self):
+        c = WhisperLiveKitConfig(backend_policy="localagreement")
+        assert c.backend_policy == "localagreement"
+
+
+class TestFromNamespace:
+    def test_known_keys(self):
+        ns = SimpleNamespace(backend="faster-whisper", lan="en", model_size="large-v3")
+        c = WhisperLiveKitConfig.from_namespace(ns)
+        assert c.backend == "faster-whisper"
+        assert c.lan == "en"
+        assert c.model_size == "large-v3"
+
+    def test_ignores_unknown_keys(self):
+        ns = SimpleNamespace(backend="auto", unknown_key="value", another="x")
+        c = WhisperLiveKitConfig.from_namespace(ns)
+        assert c.backend == "auto"
+        assert not hasattr(c, "unknown_key")
+
+    def test_preserves_defaults_for_missing(self):
+        ns = SimpleNamespace(backend="voxtral-mlx")
+        c = WhisperLiveKitConfig.from_namespace(ns)
+        assert c.lan == "auto"
+        assert c.vac is True
+
+
+class TestFromKwargs:
+    def test_known_keys(self):
+        c = WhisperLiveKitConfig.from_kwargs(backend="mlx-whisper", lan="fr")
+        assert c.backend == "mlx-whisper"
+        assert c.lan == "fr"
+
+    def test_warns_on_unknown_keys(self, caplog):
+        with caplog.at_level(logging.WARNING, logger="whisperlivekit.config"):
+            c = WhisperLiveKitConfig.from_kwargs(backend="auto", bogus="value")
+        assert c.backend == "auto"
+        assert "bogus" in caplog.text
+
+    def test_post_init_runs(self):
+        c = WhisperLiveKitConfig.from_kwargs(model_size="small.en")
+        assert c.lan == "en"

tests/test_hypothesis_buffer.py

@@ -0,0 +1,172 @@
+"""Tests for HypothesisBuffer — the core of LocalAgreement policy."""
+
+import pytest
+
+from whisperlivekit.timed_objects import ASRToken
+from whisperlivekit.local_agreement.online_asr import HypothesisBuffer
+
+
+def make_tokens(words, start=0.0, step=0.5):
+    """Helper: create ASRToken list from word strings."""
+    tokens = []
+    t = start
+    for w in words:
+        tokens.append(ASRToken(start=t, end=t + step, text=w, probability=0.9))
+        t += step
+    return tokens
+
+
+class TestInsert:
+    def test_basic_insert(self):
+        buf = HypothesisBuffer()
+        tokens = make_tokens(["hello", "world"])
+        buf.insert(tokens, offset=0.0)
+        assert len(buf.new) == 2
+        assert buf.new[0].text == "hello"
+
+    def test_insert_with_offset(self):
+        buf = HypothesisBuffer()
+        tokens = make_tokens(["hello"], start=0.0)
+        buf.insert(tokens, offset=5.0)
+        assert buf.new[0].start == pytest.approx(5.0)
+
+    def test_insert_filters_old_tokens(self):
+        buf = HypothesisBuffer()
+        buf.last_committed_time = 10.0
+        tokens = make_tokens(["old", "new"], start=5.0, step=3.0)
+        buf.insert(tokens, offset=0.0)
+        # "old" at 5.0 is before last_committed_time - 0.1 = 9.9 → filtered
+        # "new" at 8.0 is also before 9.9 → filtered
+        assert len(buf.new) == 0
+
+    def test_insert_deduplicates_committed(self):
+        buf = HypothesisBuffer()
+        # Commit "hello"
+        tokens1 = make_tokens(["hello", "world"])
+        buf.insert(tokens1, offset=0.0)
+        buf.flush()  # commits "hello" (buffer was empty, so nothing matches)
+        # Actually with empty buffer, flush won't commit anything
+        # Let's do it properly: two rounds
+        buf2 = HypothesisBuffer()
+        first = make_tokens(["hello", "world"])
+        buf2.insert(first, offset=0.0)
+        buf2.flush()  # buffer was empty → no commits, buffer = ["hello", "world"]
+
+        second = make_tokens(["hello", "world", "test"])
+        buf2.insert(second, offset=0.0)
+        committed = buf2.flush()
+        # LCP of ["hello", "world"] and ["hello", "world", "test"] = ["hello", "world"]
+        assert len(committed) == 2
+        assert committed[0].text == "hello"
+        assert committed[1].text == "world"
+
+
+class TestFlush:
+    def test_flush_empty(self):
+        buf = HypothesisBuffer()
+        committed = buf.flush()
+        assert committed == []
+
+    def test_flush_lcp_matching(self):
+        buf = HypothesisBuffer()
+        # Round 1: establish buffer
+        buf.insert(make_tokens(["hello", "world"]), offset=0.0)
+        buf.flush()  # buffer = ["hello", "world"], committed = []
+
+        # Round 2: same prefix, new suffix
+        buf.insert(make_tokens(["hello", "world", "test"]), offset=0.0)
+        committed = buf.flush()
+        assert [t.text for t in committed] == ["hello", "world"]
+
+    def test_flush_no_match(self):
+        buf = HypothesisBuffer()
+        # Round 1
+        buf.insert(make_tokens(["hello", "world"]), offset=0.0)
+        buf.flush()
+
+        # Round 2: completely different
+        buf.insert(make_tokens(["foo", "bar"]), offset=0.0)
+        committed = buf.flush()
+        assert committed == []
+
+    def test_flush_partial_match(self):
+        buf = HypothesisBuffer()
+        buf.insert(make_tokens(["hello", "world", "test"]), offset=0.0)
+        buf.flush()
+
+        buf.insert(make_tokens(["hello", "earth", "again"]), offset=0.0)
+        committed = buf.flush()
+        assert len(committed) == 1
+        assert committed[0].text == "hello"
+
+    def test_flush_updates_last_committed(self):
+        buf = HypothesisBuffer()
+        buf.insert(make_tokens(["hello", "world"]), offset=0.0)
+        buf.flush()
+
+        buf.insert(make_tokens(["hello", "world", "test"]), offset=0.0)
+        buf.flush()
+        assert buf.last_committed_word == "world"
+        assert buf.last_committed_time > 0
+
+    def test_flush_with_confidence_validation(self):
+        buf = HypothesisBuffer(confidence_validation=True)
+        high_conf = [
+            ASRToken(start=0.0, end=0.5, text="sure", probability=0.99),
+            ASRToken(start=0.5, end=1.0, text="maybe", probability=0.5),
+        ]
+        buf.insert(high_conf, offset=0.0)
+        committed = buf.flush()
+        # "sure" has p>0.95 → committed immediately
+        assert len(committed) == 1
+        assert committed[0].text == "sure"
+
+
+class TestPopCommitted:
+    def test_pop_removes_old(self):
+        buf = HypothesisBuffer()
+        buf.committed_in_buffer = make_tokens(["a", "b", "c"], start=0.0, step=1.0)
+        # "a": end=1.0, "b": end=2.0, "c": end=3.0
+        # pop_committed removes tokens with end <= time
+        buf.pop_committed(2.0)
+        # "a" (end=1.0) and "b" (end=2.0) removed, "c" (end=3.0) remains
+        assert len(buf.committed_in_buffer) == 1
+        assert buf.committed_in_buffer[0].text == "c"
+
+    def test_pop_nothing(self):
+        buf = HypothesisBuffer()
+        buf.committed_in_buffer = make_tokens(["a", "b"], start=5.0)
+        buf.pop_committed(0.0)
+        assert len(buf.committed_in_buffer) == 2
+
+    def test_pop_all(self):
+        buf = HypothesisBuffer()
+        buf.committed_in_buffer = make_tokens(["a", "b"], start=0.0, step=0.5)
+        buf.pop_committed(100.0)
+        assert len(buf.committed_in_buffer) == 0
+
+
+class TestStreamingSimulation:
+    """Multi-round insert/flush simulating real streaming behavior."""
+
+    def test_three_rounds(self):
+        buf = HypothesisBuffer()
+        all_committed = []
+
+        # Round 1: "this is"
+        buf.insert(make_tokens(["this", "is"]), offset=0.0)
+        all_committed.extend(buf.flush())
+
+        # Round 2: "this is a test"
+        buf.insert(make_tokens(["this", "is", "a", "test"]), offset=0.0)
+        all_committed.extend(buf.flush())
+
+        # Round 3: "this is a test today"
+        buf.insert(make_tokens(["this", "is", "a", "test", "today"]), offset=0.0)
+        all_committed.extend(buf.flush())
+
+        words = [t.text for t in all_committed]
+        assert "this" in words
+        assert "is" in words
+        assert "a" in words
+        assert "test" in words

tests/test_metrics.py

@@ -0,0 +1,183 @@
+"""Tests for whisperlivekit.metrics — WER, timestamp accuracy, normalization."""
+
+import pytest
+
+from whisperlivekit.metrics import compute_wer, compute_timestamp_accuracy, normalize_text
+
+
+class TestNormalizeText:
+    def test_lowercase(self):
+        assert normalize_text("Hello World") == "hello world"
+
+    def test_strip_punctuation(self):
+        assert normalize_text("Hello, world!") == "hello world"
+
+    def test_collapse_whitespace(self):
+        assert normalize_text("  hello   world  ") == "hello world"
+
+    def test_keep_hyphens(self):
+        assert normalize_text("real-time") == "real-time"
+
+    def test_keep_apostrophes(self):
+        assert normalize_text("don't") == "don't"
+
+    def test_unicode_normalized(self):
+        # e + combining accent should be same as precomposed
+        assert normalize_text("caf\u0065\u0301") == normalize_text("caf\u00e9")
+
+    def test_empty(self):
+        assert normalize_text("") == ""
+
+    def test_only_punctuation(self):
+        assert normalize_text("...!?") == ""
+
+
+class TestComputeWER:
+    def test_perfect_match(self):
+        result = compute_wer("hello world", "hello world")
+        assert result["wer"] == 0.0
+        assert result["substitutions"] == 0
+        assert result["insertions"] == 0
+        assert result["deletions"] == 0
+
+    def test_case_insensitive(self):
+        result = compute_wer("Hello World", "hello world")
+        assert result["wer"] == 0.0
+
+    def test_punctuation_ignored(self):
+        result = compute_wer("Hello, world!", "hello world")
+        assert result["wer"] == 0.0
+
+    def test_one_substitution(self):
+        result = compute_wer("hello world", "hello earth")
+        assert result["wer"] == pytest.approx(0.5)
+        assert result["substitutions"] == 1
+
+    def test_one_insertion(self):
+        result = compute_wer("hello world", "hello big world")
+        assert result["wer"] == pytest.approx(0.5)
+        assert result["insertions"] == 1
+
+    def test_one_deletion(self):
+        result = compute_wer("hello big world", "hello world")
+        assert result["wer"] == pytest.approx(1 / 3)
+        assert result["deletions"] == 1
+
+    def test_completely_different(self):
+        result = compute_wer("the cat sat", "a dog ran")
+        assert result["wer"] == pytest.approx(1.0)
+
+    def test_empty_reference(self):
+        result = compute_wer("", "hello")
+        assert result["wer"] == 1.0  # 1 insertion / 0 ref → treated as float(m)
+        assert result["ref_words"] == 0
+
+    def test_empty_hypothesis(self):
+        result = compute_wer("hello world", "")
+        assert result["wer"] == pytest.approx(1.0)
+        assert result["deletions"] == 2
+
+    def test_both_empty(self):
+        result = compute_wer("", "")
+        assert result["wer"] == 0.0
+
+    def test_ref_and_hyp_word_counts(self):
+        result = compute_wer("one two three", "one two three four")
+        assert result["ref_words"] == 3
+        assert result["hyp_words"] == 4
+
+
+class TestComputeTimestampAccuracy:
+    def test_perfect_match(self):
+        words = [
+            {"word": "hello", "start": 0.0, "end": 0.5},
+            {"word": "world", "start": 0.5, "end": 1.0},
+        ]
+        result = compute_timestamp_accuracy(words, words)
+        assert result["mae_start"] == 0.0
+        assert result["max_delta_start"] == 0.0
+        assert result["n_matched"] == 2
+
+    def test_constant_offset(self):
+        ref = [
+            {"word": "hello", "start": 0.0, "end": 0.5},
+            {"word": "world", "start": 0.5, "end": 1.0},
+        ]
+        pred = [
+            {"word": "hello", "start": 0.1, "end": 0.6},
+            {"word": "world", "start": 0.6, "end": 1.1},
+        ]
+        result = compute_timestamp_accuracy(pred, ref)
+        assert result["mae_start"] == pytest.approx(0.1)
+        assert result["max_delta_start"] == pytest.approx(0.1)
+        assert result["n_matched"] == 2
+
+    def test_mismatched_word_counts(self):
+        ref = [
+            {"word": "hello", "start": 0.0, "end": 0.5},
+            {"word": "beautiful", "start": 0.5, "end": 1.0},
+            {"word": "world", "start": 1.0, "end": 1.5},
+        ]
+        pred = [
+            {"word": "hello", "start": 0.0, "end": 0.5},
+            {"word": "world", "start": 1.1, "end": 1.6},
+        ]
+        result = compute_timestamp_accuracy(pred, ref)
+        assert result["n_matched"] == 2
+        assert result["n_ref"] == 3
+        assert result["n_pred"] == 2
+
+    def test_empty_predicted(self):
+        ref = [{"word": "hello", "start": 0.0, "end": 0.5}]
+        result = compute_timestamp_accuracy([], ref)
+        assert result["mae_start"] is None
+        assert result["n_matched"] == 0
+
+    def test_empty_reference(self):
+        pred = [{"word": "hello", "start": 0.0, "end": 0.5}]
+        result = compute_timestamp_accuracy(pred, [])
+        assert result["mae_start"] is None
+        assert result["n_matched"] == 0
+
+    def test_case_insensitive_matching(self):
+        ref = [{"word": "Hello", "start": 0.0, "end": 0.5}]
+        pred = [{"word": "hello", "start": 0.1, "end": 0.6}]
+        result = compute_timestamp_accuracy(pred, ref)
+        assert result["n_matched"] == 1
+        assert result["mae_start"] == pytest.approx(0.1)
+
+    def test_median_even_count(self):
+        """Median with even number of matched words should average the two middle values."""
+        ref = [
+            {"word": "a", "start": 0.0, "end": 0.2},
+            {"word": "b", "start": 0.5, "end": 0.7},
+            {"word": "c", "start": 1.0, "end": 1.2},
+            {"word": "d", "start": 1.5, "end": 1.7},
+        ]
+        pred = [
+            {"word": "a", "start": 0.1, "end": 0.3},   # delta 0.1
+            {"word": "b", "start": 0.7, "end": 0.9},   # delta 0.2
+            {"word": "c", "start": 1.3, "end": 1.5},   # delta 0.3
+            {"word": "d", "start": 1.9, "end": 2.1},   # delta 0.4
+        ]
+        result = compute_timestamp_accuracy(pred, ref)
+        assert result["n_matched"] == 4
+        # sorted abs deltas: [0.1, 0.2, 0.3, 0.4] -> median = (0.2 + 0.3) / 2 = 0.25
+        assert result["median_delta_start"] == pytest.approx(0.25)
+
+    def test_median_odd_count(self):
+        """Median with odd number of matched words takes the middle value."""
+        ref = [
+            {"word": "a", "start": 0.0, "end": 0.2},
+            {"word": "b", "start": 0.5, "end": 0.7},
+            {"word": "c", "start": 1.0, "end": 1.2},
+        ]
+        pred = [
+            {"word": "a", "start": 0.1, "end": 0.3},   # delta 0.1
+            {"word": "b", "start": 0.8, "end": 1.0},   # delta 0.3
+            {"word": "c", "start": 1.2, "end": 1.4},   # delta 0.2
+        ]
+        result = compute_timestamp_accuracy(pred, ref)
+        assert result["n_matched"] == 3
+        # sorted abs deltas: [0.1, 0.2, 0.3] -> median = 0.2
+        assert result["median_delta_start"] == pytest.approx(0.2)

tests/test_silence_handling.py

@@ -0,0 +1,99 @@
+"""Tests for silence handling — state machine and double-counting regression."""
+
+import pytest
+
+from whisperlivekit.timed_objects import Silence
+
+
+class TestSilenceStateMachine:
+    """Test Silence object state transitions."""
+
+    def test_initial_state(self):
+        s = Silence(start=1.0, is_starting=True)
+        assert s.is_starting is True
+        assert s.has_ended is False
+        assert s.duration is None
+        assert s.end is None
+
+    def test_end_silence(self):
+        s = Silence(start=1.0, is_starting=True)
+        s.end = 3.0
+        s.is_starting = False
+        s.has_ended = True
+        s.compute_duration()
+        assert s.duration == pytest.approx(2.0)
+
+    def test_very_short_silence(self):
+        s = Silence(start=1.0, end=1.01, is_starting=False, has_ended=True)
+        s.compute_duration()
+        assert s.duration == pytest.approx(0.01)
+
+    def test_zero_duration_silence(self):
+        s = Silence(start=5.0, end=5.0)
+        s.compute_duration()
+        assert s.duration == pytest.approx(0.0)
+
+
+class TestSilenceDoubleCounting:
+    """Regression tests for the silence double-counting bug.
+
+    The bug: _begin_silence and _end_silence both pushed self.current_silence
+    to the queue. Since they were the same Python object, _end_silence's mutation
+    affected the already-queued start event. The consumer processed both as
+    ended silences, doubling the duration.
+
+    Fix: _begin_silence now pushes a separate Silence object for the start event.
+    """
+
+    def test_start_and_end_are_separate_objects(self):
+        """Simulate the fix: start event and end event must be different objects."""
+        # Simulate _begin_silence: creates start event as separate object
+        current_silence = Silence(start=1.0, is_starting=True)
+        start_event = Silence(start=1.0, is_starting=True)  # separate copy
+
+        # Simulate _end_silence: mutates current_silence
+        current_silence.end = 3.0
+        current_silence.is_starting = False
+        current_silence.has_ended = True
+        current_silence.compute_duration()
+
+        # start_event should NOT be affected by mutations to current_silence
+        assert start_event.is_starting is True
+        assert start_event.has_ended is False
+        assert start_event.end is None
+
+        # current_silence (end event) has the final state
+        assert current_silence.has_ended is True
+        assert current_silence.duration == pytest.approx(2.0)
+
+    def test_single_object_would_cause_double_counting(self):
+        """Demonstrate the bug: if same object is used for both events."""
+        shared = Silence(start=1.0, is_starting=True)
+        queue = [shared]  # start event queued
+
+        # Mutate (simulates _end_silence)
+        shared.end = 3.0
+        shared.is_starting = False
+        shared.has_ended = True
+        shared.compute_duration()
+        queue.append(shared)  # end event queued
+
+        # Both queue items point to the SAME mutated object
+        assert queue[0] is queue[1]  # same reference
+        assert queue[0].has_ended is True  # start event also shows ended!
+
+        # This would cause double-counting: both items have has_ended=True
+        # and duration=2.0, so the consumer adds 2.0 twice = 4.0
+
+
+class TestConsecutiveSilences:
+    def test_multiple_silences(self):
+        """Multiple silence periods should have independent durations."""
+        s1 = Silence(start=1.0, end=2.0)
+        s1.compute_duration()
+        s2 = Silence(start=5.0, end=8.0)
+        s2.compute_duration()
+        assert s1.duration == pytest.approx(1.0)
+        assert s2.duration == pytest.approx(3.0)
+        # Total silence should be sum, not accumulated on single object
+        assert s1.duration + s2.duration == pytest.approx(4.0)

tests/test_timed_objects.py

@@ -0,0 +1,185 @@
+"""Tests for whisperlivekit.timed_objects data classes."""
+
+import pytest
+
+from whisperlivekit.timed_objects import (
+    ASRToken,
+    FrontData,
+    Segment,
+    Silence,
+    TimedText,
+    Transcript,
+    format_time,
+)
+
+
+class TestFormatTime:
+    def test_zero(self):
+        assert format_time(0) == "0:00:00"
+
+    def test_one_minute(self):
+        assert format_time(60) == "0:01:00"
+
+    def test_one_hour(self):
+        assert format_time(3600) == "1:00:00"
+
+    def test_fractional_truncated(self):
+        assert format_time(61.9) == "0:01:01"
+
+
+class TestASRToken:
+    def test_with_offset(self):
+        t = ASRToken(start=1.0, end=2.0, text="hello")
+        shifted = t.with_offset(0.5)
+        assert shifted.start == pytest.approx(1.5)
+        assert shifted.end == pytest.approx(2.5)
+        assert shifted.text == "hello"
+
+    def test_with_offset_preserves_fields(self):
+        t = ASRToken(start=0.0, end=1.0, text="hi", speaker=2, probability=0.95)
+        shifted = t.with_offset(1.0)
+        assert shifted.speaker == 2
+        assert shifted.probability == 0.95
+
+    def test_is_silence_false(self):
+        t = ASRToken(start=0.0, end=1.0, text="hello")
+        assert t.is_silence() is False
+
+    def test_bool_truthy(self):
+        t = ASRToken(start=0.0, end=1.0, text="hello")
+        assert bool(t) is True
+
+    def test_bool_falsy(self):
+        t = ASRToken(start=0.0, end=1.0, text="")
+        assert bool(t) is False
+
+
+class TestTimedText:
+    def test_has_punctuation_period(self):
+        t = TimedText(text="hello.")
+        assert t.has_punctuation() is True
+
+    def test_has_punctuation_exclamation(self):
+        t = TimedText(text="wow!")
+        assert t.has_punctuation() is True
+
+    def test_has_punctuation_question(self):
+        t = TimedText(text="really?")
+        assert t.has_punctuation() is True
+
+    def test_has_punctuation_cjk(self):
+        t = TimedText(text="hello。")
+        assert t.has_punctuation() is True
+
+    def test_no_punctuation(self):
+        t = TimedText(text="hello world")
+        assert t.has_punctuation() is False
+
+    def test_duration(self):
+        t = TimedText(start=1.0, end=3.5)
+        assert t.duration() == pytest.approx(2.5)
+
+    def test_contains_timespan(self):
+        outer = TimedText(start=0.0, end=5.0)
+        inner = TimedText(start=1.0, end=3.0)
+        assert outer.contains_timespan(inner) is True
+        assert inner.contains_timespan(outer) is False
+
+
+class TestSilence:
+    def test_compute_duration(self):
+        s = Silence(start=1.0, end=3.5)
+        d = s.compute_duration()
+        assert d == pytest.approx(2.5)
+        assert s.duration == pytest.approx(2.5)
+
+    def test_compute_duration_none_start(self):
+        s = Silence(start=None, end=3.5)
+        d = s.compute_duration()
+        assert d is None
+
+    def test_compute_duration_none_end(self):
+        s = Silence(start=1.0, end=None)
+        d = s.compute_duration()
+        assert d is None
+
+    def test_is_silence_true(self):
+        s = Silence()
+        assert s.is_silence() is True
+
+
+class TestTranscript:
+    def test_from_tokens(self, sample_tokens):
+        t = Transcript.from_tokens(sample_tokens, sep="")
+        assert t.text == "Hello world test."
+        assert t.start == pytest.approx(0.0)
+        assert t.end == pytest.approx(1.5)
+
+    def test_from_tokens_with_sep(self, sample_tokens):
+        t = Transcript.from_tokens(sample_tokens, sep="|")
+        assert t.text == "Hello| world| test."
+
+    def test_from_empty_tokens(self):
+        t = Transcript.from_tokens([])
+        assert t.text == ""
+        assert t.start is None
+        assert t.end is None
+
+    def test_from_tokens_with_offset(self, sample_tokens):
+        t = Transcript.from_tokens(sample_tokens, offset=10.0)
+        assert t.start == pytest.approx(10.0)
+        assert t.end == pytest.approx(11.5)
+
+
+class TestSegment:
+    def test_from_tokens(self, sample_tokens):
+        seg = Segment.from_tokens(sample_tokens)
+        assert seg is not None
+        assert seg.text == "Hello world test."
+        assert seg.start == pytest.approx(0.0)
+        assert seg.end == pytest.approx(1.5)
+        assert seg.speaker == -1
+
+    def test_from_silence_tokens(self):
+        silences = [
+            Silence(start=1.0, end=2.0),
+            Silence(start=2.0, end=3.0),
+        ]
+        seg = Segment.from_tokens(silences, is_silence=True)
+        assert seg is not None
+        assert seg.speaker == -2
+        assert seg.is_silence() is True
+        assert seg.text is None
+
+    def test_from_empty_tokens(self):
+        seg = Segment.from_tokens([])
+        assert seg is None
+
+    def test_to_dict(self, sample_tokens):
+        seg = Segment.from_tokens(sample_tokens)
+        d = seg.to_dict()
+        assert "text" in d
+        assert "speaker" in d
+        assert "start" in d
+        assert "end" in d
+
+
+class TestFrontData:
+    def test_to_dict_empty(self):
+        fd = FrontData()
+        d = fd.to_dict()
+        assert d["lines"] == []
+        assert d["buffer_transcription"] == ""
+        assert "error" not in d
+
+    def test_to_dict_with_error(self):
+        fd = FrontData(error="something broke")
+        d = fd.to_dict()
+        assert d["error"] == "something broke"
+
+    def test_to_dict_with_lines(self, sample_tokens):
+        seg = Segment.from_tokens(sample_tokens)
+        fd = FrontData(lines=[seg])
+        d = fd.to_dict()
+        assert len(d["lines"]) == 1
+        assert d["lines"][0]["text"] == "Hello world test."

whisperlivekit/__init__.py

@@ -1,7 +1,7 @@
 from .audio_processor import AudioProcessor
 from .core import TranscriptionEngine
 from .parse_args import parse_args
-from .web.web_interface import get_web_interface_html, get_inline_ui_html
+from .web.web_interface import get_inline_ui_html, get_web_interface_html
 
 __all__ = [
     "TranscriptionEngine",

whisperlivekit/audio_processor.py

@@ -1,14 +1,21 @@
 import asyncio
-import numpy as np
-from time import time
 import logging
 import traceback
-from typing import Optional, Union, List, Any, AsyncGenerator
-from whisperlivekit.timed_objects import ASRToken, Silence, Line, FrontData, State, Transcript, ChangeSpeaker
-from whisperlivekit.core import TranscriptionEngine, online_factory, online_diarization_factory, online_translation_factory
-from whisperlivekit.silero_vad_iterator import FixedVADIterator
+from time import time
+from typing import Any, AsyncGenerator, List, Optional, Union
+
+import numpy as np
+
+from whisperlivekit.core import (TranscriptionEngine,
+                                 online_diarization_factory, online_factory,
+                                 online_translation_factory)
+from whisperlivekit.metrics_collector import SessionMetrics
 from whisperlivekit.ffmpeg_manager import FFmpegManager, FFmpegState
+from whisperlivekit.silero_vad_iterator import FixedVADIterator, OnnxWrapper, load_jit_vad
+from whisperlivekit.timed_objects import (ASRToken, ChangeSpeaker, FrontData,
+                                          Segment, Silence, State, Transcript)
 from whisperlivekit.tokens_alignment import TokensAlignment
+
 logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
 logger = logging.getLogger(__name__)
 logger.setLevel(logging.DEBUG)
@@ -26,7 +33,7 @@ async def get_all_from_queue(queue: asyncio.Queue) -> Union[object, Silence, np.
     if isinstance(first_item, Silence):
         return first_item
     items.append(first_item)
-    
+
     while True:
         if not queue._queue:
             break
@@ -47,20 +54,21 @@ class AudioProcessor:
     Processes audio streams for transcription and diarization.
     Handles audio processing, state management, and result formatting.
     """
-    
+
     def __init__(self, **kwargs: Any) -> None:
         """Initialize the audio processor with configuration, models, and state."""
-        
+
         if 'transcription_engine' in kwargs and isinstance(kwargs['transcription_engine'], TranscriptionEngine):
             models = kwargs['transcription_engine']
         else:
             models = TranscriptionEngine(**kwargs)
-        
+
         # Audio processing settings
         self.args = models.args
         self.sample_rate = 16000
         self.channels = 1
-        self.samples_per_sec = int(self.sample_rate * self.args.min_chunk_size)
+        chunk_seconds = self.args.vac_chunk_size if self.args.vac else self.args.min_chunk_size
+        self.samples_per_sec = int(self.sample_rate * chunk_seconds)
         self.bytes_per_sample = 2
         self.bytes_per_sec = self.samples_per_sec * self.bytes_per_sample
         self.max_bytes_per_sec = 32000 * 5  # 5 seconds of audio at 32 kHz
@@ -79,12 +87,14 @@ class AudioProcessor:
 
         # Models and processing
         self.asr: Any = models.asr
-        self.vac_model: Any = models.vac_model
+        self.vac: Optional[FixedVADIterator] = None
+
         if self.args.vac:
-            self.vac: Optional[FixedVADIterator] = FixedVADIterator(models.vac_model)
-        else:
-            self.vac: Optional[FixedVADIterator] = None
-                         
+            if models.vac_session is not None:
+                vac_model = OnnxWrapper(session=models.vac_session)
+                self.vac = FixedVADIterator(vac_model)
+            else:
+                self.vac = FixedVADIterator(load_jit_vad())
         self.ffmpeg_manager: Optional[FFmpegManager] = None
         self.ffmpeg_reader_task: Optional[asyncio.Task] = None
         self._ffmpeg_error: Optional[str] = None
@@ -98,7 +108,7 @@ class AudioProcessor:
                 logger.error(f"FFmpeg error: {error_type}")
                 self._ffmpeg_error = error_type
             self.ffmpeg_manager.on_error_callback = handle_ffmpeg_error
-             
+
         self.transcription_queue: Optional[asyncio.Queue] = asyncio.Queue() if self.args.transcription else None
         self.diarization_queue: Optional[asyncio.Queue] = asyncio.Queue() if self.args.diarization else None
         self.translation_queue: Optional[asyncio.Queue] = asyncio.Queue() if self.args.target_language else None
@@ -109,14 +119,15 @@ class AudioProcessor:
         self.translation_task: Optional[asyncio.Task] = None
         self.watchdog_task: Optional[asyncio.Task] = None
         self.all_tasks_for_cleanup: List[asyncio.Task] = []
-        
+        self.metrics: SessionMetrics = SessionMetrics()
+
         self.transcription: Optional[Any] = None
         self.translation: Optional[Any] = None
         self.diarization: Optional[Any] = None
 
         if self.args.transcription:
-            self.transcription = online_factory(self.args, models.asr)        
-            self.sep = self.transcription.asr.sep   
+            self.transcription = online_factory(self.args, models.asr)
+            self.sep = self.transcription.asr.sep
         if self.args.diarization:
             self.diarization = online_diarization_factory(self.args, models.diarization_model)
         if models.translation_model:
@@ -130,25 +141,43 @@ class AudioProcessor:
         if self.translation_queue:
             await self.translation_queue.put(self.current_silence)
 
-    async def _begin_silence(self) -> None:
+    async def _begin_silence(self, at_sample: Optional[int] = None) -> None:
         if self.current_silence:
             return
-        now = time() - self.beg_loop
+        # Use audio stream time (sample-precise) for accurate silence duration
+        if at_sample is not None:
+            audio_t = at_sample / self.sample_rate
+        else:
+            audio_t = self.total_pcm_samples / self.sample_rate if self.sample_rate else 0.0
         self.current_silence = Silence(
-            is_starting=True, start=now
+            is_starting=True, start=audio_t
         )
-        await self._push_silence_event()
+        # Push a separate start-only event so _end_silence won't mutate it
+        start_event = Silence(is_starting=True, start=audio_t)
+        if self.transcription_queue:
+            await self.transcription_queue.put(start_event)
+        if self.args.diarization and self.diarization_queue:
+            await self.diarization_queue.put(start_event)
+        if self.translation_queue:
+            await self.translation_queue.put(start_event)
 
-    async def _end_silence(self) -> None:
+    async def _end_silence(self, at_sample: Optional[int] = None) -> None:
         if not self.current_silence:
             return
-        now = time() - self.beg_loop
-        self.current_silence.end = now
-        self.current_silence.is_starting=False
-        self.current_silence.has_ended=True
+        if at_sample is not None:
+            audio_t = at_sample / self.sample_rate
+        else:
+            audio_t = self.total_pcm_samples / self.sample_rate if self.sample_rate else 0.0
+        self.current_silence.end = audio_t
+        self.current_silence.is_starting = False
+        self.current_silence.has_ended = True
         self.current_silence.compute_duration()
+        self.metrics.n_silence_events += 1
+        if self.current_silence.duration is not None:
+            self.metrics.total_silence_duration_s += self.current_silence.duration
         if self.current_silence.duration > MIN_DURATION_REAL_SILENCE:
             self.state.new_tokens.append(self.current_silence)
+        # Push the completed silence as the end event (separate from the start event)
         await self._push_silence_event()
         self.current_silence = None
 
@@ -174,24 +203,24 @@ class AudioProcessor:
     def convert_pcm_to_float(self, pcm_buffer: Union[bytes, bytearray]) -> np.ndarray:
         """Convert PCM buffer in s16le format to normalized NumPy array."""
         return np.frombuffer(pcm_buffer, dtype=np.int16).astype(np.float32) / 32768.0
-            
+
     async def get_current_state(self) -> State:
         """Get current state."""
         async with self.lock:
             current_time = time()
-            
+
             remaining_transcription = 0
             if self.state.end_buffer > 0:
                 remaining_transcription = max(0, round(current_time - self.beg_loop - self.state.end_buffer, 1))
-                
+
             remaining_diarization = 0
             if self.state.tokens:
                 latest_end = max(self.state.end_buffer, self.state.tokens[-1].end if self.state.tokens else 0)
                 remaining_diarization = max(0, round(latest_end - self.state.end_attributed_speaker, 1))
-                
+
             self.state.remaining_time_transcription = remaining_transcription
             self.state.remaining_time_diarization = remaining_diarization
-            
+
             return self.state
 
     async def ffmpeg_stdout_reader(self) -> None:
@@ -244,16 +273,45 @@ class AudioProcessor:
         if self.translation:
             await self.translation_queue.put(SENTINEL)
 
+    async def _finish_transcription(self) -> None:
+        """Call finish() on the online processor to flush remaining tokens."""
+        if not self.transcription:
+            return
+        try:
+            if hasattr(self.transcription, 'finish'):
+                final_tokens, end_time = await asyncio.to_thread(self.transcription.finish)
+            else:
+                # SimulStreamingOnlineProcessor uses start_silence() → process_iter(is_last=True)
+                final_tokens, end_time = await asyncio.to_thread(self.transcription.start_silence)
+
+            final_tokens = final_tokens or []
+            if final_tokens:
+                logger.info(f"Finish flushed {len(final_tokens)} tokens")
+                _buffer_transcript = self.transcription.get_buffer()
+                async with self.lock:
+                    self.state.tokens.extend(final_tokens)
+                    self.state.buffer_transcription = _buffer_transcript
+                    self.state.end_buffer = max(self.state.end_buffer, end_time)
+                    self.state.new_tokens.extend(final_tokens)
+                    self.state.new_tokens_buffer = _buffer_transcript
+                if self.translation_queue:
+                    for token in final_tokens:
+                        await self.translation_queue.put(token)
+        except Exception as e:
+            logger.warning(f"Error finishing transcription: {e}")
+            logger.debug(f"Traceback: {traceback.format_exc()}")
+
     async def transcription_processor(self) -> None:
         """Process audio chunks for transcription."""
         cumulative_pcm_duration_stream_time = 0.0
-        
+
         while True:
             try:
                 # item = await self.transcription_queue.get()
                 item = await get_all_from_queue(self.transcription_queue)
                 if item is SENTINEL:
                     logger.debug("Transcription processor received sentinel. Finishing.")
+                    await self._finish_transcription()
                     break
 
                 asr_internal_buffer_duration_s = len(getattr(self.transcription, 'audio_buffer', [])) / self.transcription.SAMPLING_RATE
@@ -288,8 +346,13 @@ class AudioProcessor:
                     cumulative_pcm_duration_stream_time += len(pcm_array) / self.sample_rate
                     stream_time_end_of_current_pcm = cumulative_pcm_duration_stream_time
                     self.transcription.insert_audio_chunk(pcm_array, stream_time_end_of_current_pcm)
+                    _t0 = time()
                     new_tokens, current_audio_processed_upto = await asyncio.to_thread(self.transcription.process_iter)
+                    _dur = time() - _t0
+                    self.metrics.transcription_durations.append(_dur)
+                    self.metrics.n_transcription_calls += 1
                     new_tokens = new_tokens or []
+                    self.metrics.n_tokens_produced += len(new_tokens)
 
                 _buffer_transcript = self.transcription.get_buffer()
                 buffer_text = _buffer_transcript.text
@@ -303,12 +366,12 @@ class AudioProcessor:
 
                 if new_tokens:
                     candidate_end_times.append(new_tokens[-1].end)
-                
+
                 if _buffer_transcript.end is not None:
                     candidate_end_times.append(_buffer_transcript.end)
-                
+
                 candidate_end_times.append(current_audio_processed_upto)
-                
+
                 async with self.lock:
                     self.state.tokens.extend(new_tokens)
                     self.state.buffer_transcription = _buffer_transcript
@@ -318,13 +381,13 @@ class AudioProcessor:
 
                 if self.translation_queue:
                     for token in new_tokens:
-                        await self.translation_queue.put(token)                
+                        await self.translation_queue.put(token)
             except Exception as e:
                 logger.warning(f"Exception in transcription_processor: {e}")
                 logger.warning(f"Traceback: {traceback.format_exc()}")
                 if 'pcm_array' in locals() and pcm_array is not SENTINEL : # Check if pcm_array was assigned from queue
                     self.transcription_queue.task_done()
-        
+
         if self.is_stopping:
             logger.info("Transcription processor finishing due to stopping flag.")
             if self.diarization_queue:
@@ -345,18 +408,21 @@ class AudioProcessor:
                     if item.has_ended:
                         self.diarization.insert_silence(item.duration)
                     continue
-
                 self.diarization.insert_audio_chunk(item)
                 diarization_segments = await self.diarization.diarize()
-                self.state.new_diarization = diarization_segments
-                
+                diar_end = 0.0
+                if diarization_segments:
+                    diar_end = max(getattr(s, "end", 0.0) for s in diarization_segments)
+                async with self.lock:
+                    self.state.new_diarization = diarization_segments
+                    self.state.end_attributed_speaker = max(self.state.end_attributed_speaker, diar_end)
             except Exception as e:
                 logger.warning(f"Exception in diarization_processor: {e}")
                 logger.warning(f"Traceback: {traceback.format_exc()}")
         logger.info("Diarization processor task finished.")
 
     async def translation_processor(self) -> None:
-        # the idea is to ignore diarization for the moment. We use only transcription tokens. 
+        # the idea is to ignore diarization for the moment. We use only transcription tokens.
         # And the speaker is attributed given the segments used for the translation
         # in the future we want to have different languages for each speaker etc, so it will be more complex.
         while True:
@@ -418,22 +484,23 @@ class AudioProcessor:
                     remaining_time_transcription=state.remaining_time_transcription,
                     remaining_time_diarization=state.remaining_time_diarization if self.args.diarization else 0
                 )
-                                
+
                 should_push = (response != self.last_response_content)
                 if should_push:
+                    self.metrics.n_responses_sent += 1
                     yield response
                     self.last_response_content = response
-                
+
                 if self.is_stopping and self._processing_tasks_done():
                     logger.info("Results formatter: All upstream processors are done and in stopping state. Terminating.")
                     return
-                
+
                 await asyncio.sleep(0.05)
-                
+
             except Exception as e:
                 logger.warning(f"Exception in results_formatter. Traceback: {traceback.format_exc()}")
                 await asyncio.sleep(0.5)
-        
+
     async def create_tasks(self) -> AsyncGenerator[FrontData, None]:
         """Create and start processing tasks."""
         self.all_tasks_for_cleanup = []
@@ -458,21 +525,21 @@ class AudioProcessor:
             self.transcription_task = asyncio.create_task(self.transcription_processor())
             self.all_tasks_for_cleanup.append(self.transcription_task)
             processing_tasks_for_watchdog.append(self.transcription_task)
-            
+
         if self.diarization:
             self.diarization_task = asyncio.create_task(self.diarization_processor())
             self.all_tasks_for_cleanup.append(self.diarization_task)
             processing_tasks_for_watchdog.append(self.diarization_task)
-        
+
         if self.translation:
             self.translation_task = asyncio.create_task(self.translation_processor())
             self.all_tasks_for_cleanup.append(self.translation_task)
             processing_tasks_for_watchdog.append(self.translation_task)
-        
+
         # Monitor overall system health
         self.watchdog_task = asyncio.create_task(self.watchdog(processing_tasks_for_watchdog))
         self.all_tasks_for_cleanup.append(self.watchdog_task)
-        
+
         return self.results_formatter()
 
     async def watchdog(self, tasks_to_monitor: List[asyncio.Task]) -> None:
@@ -485,7 +552,7 @@ class AudioProcessor:
                     return
 
                 await asyncio.sleep(10)
-                
+
                 for i, task in enumerate(list(tasks_remaining)):
                     if task.done():
                         exc = task.exception()
@@ -495,13 +562,13 @@ class AudioProcessor:
                         else:
                             logger.info(f"{task_name} completed normally.")
                         tasks_remaining.remove(task)
-                    
+
             except asyncio.CancelledError:
                 logger.info("Watchdog task cancelled.")
                 break
             except Exception as e:
                 logger.error(f"Error in watchdog task: {e}", exc_info=True)
-        
+
     async def cleanup(self) -> None:
         """Clean up resources when processing is complete."""
         logger.info("Starting cleanup of AudioProcessor resources.")
@@ -509,7 +576,7 @@ class AudioProcessor:
         for task in self.all_tasks_for_cleanup:
             if task and not task.done():
                 task.cancel()
-            
+
         created_tasks = [t for t in self.all_tasks_for_cleanup if t]
         if created_tasks:
             await asyncio.gather(*created_tasks, return_exceptions=True)
@@ -523,6 +590,10 @@ class AudioProcessor:
                 logger.warning(f"Error stopping FFmpeg manager: {e}")
         if self.diarization:
             self.diarization.close()
+
+        # Finalize session metrics
+        self.metrics.total_audio_duration_s = self.total_pcm_samples / self.sample_rate
+        self.metrics.log_summary()
         logger.info("AudioProcessor cleanup complete.")
 
     def _processing_tasks_done(self) -> bool:
@@ -541,13 +612,18 @@ class AudioProcessor:
 
         if not self.beg_loop:
             self.beg_loop = time()
+            self.metrics.session_start = self.beg_loop
             self.current_silence = Silence(start=0.0, is_starting=True)
             self.tokens_alignment.beg_loop = self.beg_loop
 
         if not message:
             logger.info("Empty audio message received, initiating stop sequence.")
             self.is_stopping = True
-             
+
+            # Flush any remaining PCM data before signaling end-of-stream
+            if self.is_pcm_input and self.pcm_buffer:
+                await self._flush_remaining_pcm()
+
             if self.transcription_queue:
                 await self.transcription_queue.put(SENTINEL)
 
@@ -560,6 +636,8 @@ class AudioProcessor:
             logger.warning("AudioProcessor is stopping. Ignoring incoming audio.")
             return
 
+        self.metrics.n_chunks_received += 1
+
         if self.is_pcm_input:
             self.pcm_buffer.extend(message)
             await self.handle_pcm_data()
@@ -576,6 +654,11 @@ class AudioProcessor:
                     logger.warning("Failed to write audio data to FFmpeg")
 
     async def handle_pcm_data(self) -> None:
+        # Without VAC, there's no speech detector to end the initial silence.
+        # Clear it on the first audio chunk so audio actually gets enqueued.
+        if not self.args.vac and self.current_silence:
+            await self._end_silence()
+
         # Process when enough data
         if len(self.pcm_buffer) < self.bytes_per_sec:
             return
@@ -588,7 +671,7 @@ class AudioProcessor:
 
         chunk_size = min(len(self.pcm_buffer), self.max_bytes_per_sec)
         aligned_chunk_size = (chunk_size // self.bytes_per_sample) * self.bytes_per_sample
-        
+
         if aligned_chunk_size == 0:
             return
         pcm_array = self.convert_pcm_to_float(self.pcm_buffer[:aligned_chunk_size])
@@ -603,16 +686,16 @@ class AudioProcessor:
             res = self.vac(pcm_array)
 
         if res is not None:
-            silence_detected = res.get("end", 0) > res.get("start", 0)
-            if silence_detected and not self.current_silence:
+            if "start" in res and self.current_silence:
+                await self._end_silence(at_sample=res.get("start"))
+
+            if "end" in res and not self.current_silence:
                 pre_silence_chunk = self._slice_before_silence(
                     pcm_array, chunk_sample_start, res.get("end")
                 )
                 if pre_silence_chunk is not None and pre_silence_chunk.size > 0:
                     await self._enqueue_active_audio(pre_silence_chunk)
-                await self._begin_silence()
-            elif self.current_silence:
-                await self._end_silence()
+                await self._begin_silence(at_sample=res.get("end"))
 
         if not self.current_silence:
             await self._enqueue_active_audio(pcm_array)
@@ -621,3 +704,21 @@ class AudioProcessor:
 
         if not self.args.transcription and not self.args.diarization:
             await asyncio.sleep(0.1)
+
+    async def _flush_remaining_pcm(self) -> None:
+        """Flush whatever PCM data remains in the buffer, regardless of size threshold."""
+        if not self.pcm_buffer:
+            return
+        aligned_size = (len(self.pcm_buffer) // self.bytes_per_sample) * self.bytes_per_sample
+        if aligned_size == 0:
+            return
+        pcm_array = self.convert_pcm_to_float(self.pcm_buffer[:aligned_size])
+        self.pcm_buffer = self.pcm_buffer[aligned_size:]
+
+        # End any active silence so the audio gets enqueued
+        if self.current_silence:
+            await self._end_silence(at_sample=self.total_pcm_samples)
+
+        await self._enqueue_active_audio(pcm_array)
+        self.total_pcm_samples += len(pcm_array)
+        logger.info(f"Flushed remaining PCM buffer: {len(pcm_array)} samples ({len(pcm_array)/self.sample_rate:.2f}s)")

whisperlivekit/backend_support.py

@@ -29,6 +29,12 @@ def mlx_backend_available(warn_on_missing = False):
     return available
 
 
+def voxtral_hf_backend_available():
+    """Return True if HF Transformers Voxtral backend is available."""
+    return module_available("transformers")
+
+
+
 def faster_backend_available(warn_on_missing = False):
     available = module_available("faster_whisper")
     if not available and warn_on_missing and platform.system() != "Darwin":

whisperlivekit/basic_server.py

@@ -1,25 +1,26 @@
-from contextlib import asynccontextmanager
-from fastapi import FastAPI, WebSocket, WebSocketDisconnect
-from fastapi.responses import HTMLResponse
-from fastapi.middleware.cors import CORSMiddleware
-from whisperlivekit import TranscriptionEngine, AudioProcessor, get_inline_ui_html, parse_args
 import asyncio
 import logging
+from contextlib import asynccontextmanager
+
+from fastapi import FastAPI, WebSocket, WebSocketDisconnect
+from fastapi.middleware.cors import CORSMiddleware
+from fastapi.responses import HTMLResponse
+
+from whisperlivekit import (AudioProcessor, TranscriptionEngine,
+                            get_inline_ui_html, parse_args)
 
 logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
 logging.getLogger().setLevel(logging.WARNING)
 logger = logging.getLogger(__name__)
 logger.setLevel(logging.DEBUG)
 
-args = parse_args()
+config = parse_args()
 transcription_engine = None
 
 @asynccontextmanager
-async def lifespan(app: FastAPI):    
+async def lifespan(app: FastAPI):
     global transcription_engine
-    transcription_engine = TranscriptionEngine(
-        **vars(args),
-    )
+    transcription_engine = TranscriptionEngine(config=config)
     yield
 
 app = FastAPI(lifespan=lifespan)
@@ -60,7 +61,7 @@ async def websocket_endpoint(websocket: WebSocket):
     logger.info("WebSocket connection opened.")
 
     try:
-        await websocket.send_json({"type": "config", "useAudioWorklet": bool(args.pcm_input)})
+        await websocket.send_json({"type": "config", "useAudioWorklet": bool(config.pcm_input)})
     except Exception as e:
         logger.warning(f"Failed to send config to client: {e}")
             
@@ -100,26 +101,26 @@ def main():
     
     uvicorn_kwargs = {
         "app": "whisperlivekit.basic_server:app",
-        "host":args.host, 
-        "port":args.port, 
+        "host": config.host,
+        "port": config.port,
         "reload": False,
         "log_level": "info",
         "lifespan": "on",
     }
-    
+
     ssl_kwargs = {}
-    if args.ssl_certfile or args.ssl_keyfile:
-        if not (args.ssl_certfile and args.ssl_keyfile):
+    if config.ssl_certfile or config.ssl_keyfile:
+        if not (config.ssl_certfile and config.ssl_keyfile):
             raise ValueError("Both --ssl-certfile and --ssl-keyfile must be specified together.")
         ssl_kwargs = {
-            "ssl_certfile": args.ssl_certfile,
-            "ssl_keyfile": args.ssl_keyfile
+            "ssl_certfile": config.ssl_certfile,
+            "ssl_keyfile": config.ssl_keyfile,
         }
 
     if ssl_kwargs:
         uvicorn_kwargs = {**uvicorn_kwargs, **ssl_kwargs}
-    if args.forwarded_allow_ips:
-        uvicorn_kwargs = { **uvicorn_kwargs, "forwarded_allow_ips" : args.forwarded_allow_ips }
+    if config.forwarded_allow_ips:
+        uvicorn_kwargs = {**uvicorn_kwargs, "forwarded_allow_ips": config.forwarded_allow_ips}
 
     uvicorn.run(**uvicorn_kwargs)
 

whisperlivekit/config.py

@@ -0,0 +1,102 @@
+"""Typed configuration for the WhisperLiveKit pipeline."""
+import logging
+from dataclasses import dataclass, field, fields
+from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class WhisperLiveKitConfig:
+    """Single source of truth for all WhisperLiveKit configuration.
+
+    Replaces the previous dict-based parameter system in TranscriptionEngine.
+    All fields have defaults matching the prior behaviour.
+    """
+
+    # Server / global
+    host: str = "localhost"
+    port: int = 8000
+    diarization: bool = False
+    punctuation_split: bool = False
+    target_language: str = ""
+    vac: bool = True
+    vac_chunk_size: float = 0.04
+    log_level: str = "DEBUG"
+    ssl_certfile: Optional[str] = None
+    ssl_keyfile: Optional[str] = None
+    forwarded_allow_ips: Optional[str] = None
+    transcription: bool = True
+    vad: bool = True
+    pcm_input: bool = False
+    disable_punctuation_split: bool = False
+    diarization_backend: str = "sortformer"
+    backend_policy: str = "simulstreaming"
+    backend: str = "auto"
+
+    # Transcription common
+    warmup_file: Optional[str] = None
+    min_chunk_size: float = 0.1
+    model_size: str = "base"
+    model_cache_dir: Optional[str] = None
+    model_dir: Optional[str] = None
+    model_path: Optional[str] = None
+    lora_path: Optional[str] = None
+    lan: str = "auto"
+    direct_english_translation: bool = False
+
+    # LocalAgreement-specific
+    buffer_trimming: str = "segment"
+    confidence_validation: bool = False
+    buffer_trimming_sec: float = 15.0
+
+    # SimulStreaming-specific
+    disable_fast_encoder: bool = False
+    custom_alignment_heads: Optional[str] = None
+    frame_threshold: int = 25
+    beams: int = 1
+    decoder_type: Optional[str] = None
+    audio_max_len: float = 20.0
+    audio_min_len: float = 0.0
+    cif_ckpt_path: Optional[str] = None
+    never_fire: bool = False
+    init_prompt: Optional[str] = None
+    static_init_prompt: Optional[str] = None
+    max_context_tokens: Optional[int] = None
+
+    # Diarization (diart)
+    segmentation_model: str = "pyannote/segmentation-3.0"
+    embedding_model: str = "pyannote/embedding"
+
+    # Translation
+    nllb_backend: str = "transformers"
+    nllb_size: str = "600M"
+
+    def __post_init__(self):
+        # .en model suffix forces English
+        if self.model_size and self.model_size.endswith(".en"):
+            self.lan = "en"
+        # Normalize backend_policy aliases
+        if self.backend_policy == "1":
+            self.backend_policy = "simulstreaming"
+        elif self.backend_policy == "2":
+            self.backend_policy = "localagreement"
+
+    # ------------------------------------------------------------------
+    # Factory helpers
+    # ------------------------------------------------------------------
+
+    @classmethod
+    def from_namespace(cls, ns) -> "WhisperLiveKitConfig":
+        """Create config from an argparse Namespace, ignoring unknown keys."""
+        known = {f.name for f in fields(cls)}
+        return cls(**{k: v for k, v in vars(ns).items() if k in known})
+
+    @classmethod
+    def from_kwargs(cls, **kwargs) -> "WhisperLiveKitConfig":
+        """Create config from keyword arguments; warns on unknown keys."""
+        known = {f.name for f in fields(cls)}
+        unknown = set(kwargs.keys()) - known
+        if unknown:
+            logger.warning("Unknown config keys ignored: %s", unknown)
+        return cls(**{k: v for k, v in kwargs.items() if k in known})

whisperlivekit/core.py

@@ -1,130 +1,142 @@
+import logging
+import sys
+import threading
+from argparse import Namespace
+from dataclasses import asdict
+
+from whisperlivekit.config import WhisperLiveKitConfig
+from whisperlivekit.local_agreement.online_asr import OnlineASRProcessor
 from whisperlivekit.local_agreement.whisper_online import backend_factory
 from whisperlivekit.simul_whisper import SimulStreamingASR
-from whisperlivekit.local_agreement.online_asr import OnlineASRProcessor
-from argparse import Namespace
-import sys
-import logging
-
-def update_with_kwargs(_dict, kwargs):
-    _dict.update({
-        k: v for k, v in kwargs.items() if k in _dict
-    })
-    return _dict
-
 
 logger = logging.getLogger(__name__)
 
 class TranscriptionEngine:
     _instance = None
     _initialized = False
+    _lock = threading.Lock()  # Thread-safe singleton lock
     
     def __new__(cls, *args, **kwargs):
+        # Double-checked locking pattern for thread-safe singleton
         if cls._instance is None:
-            cls._instance = super().__new__(cls)
+            with cls._lock:
+                # Check again inside lock to prevent race condition
+                if cls._instance is None:
+                    cls._instance = super().__new__(cls)
         return cls._instance
     
-    def __init__(self, **kwargs):
-        if TranscriptionEngine._initialized:
-            return
+    def __init__(self, config=None, **kwargs):
+        # Thread-safe initialization check
+        with TranscriptionEngine._lock:
+            if TranscriptionEngine._initialized:
+                return
 
-        global_params = {
-            "host": "localhost",
-            "port": 8000,
-            "diarization": False,
-            "punctuation_split": False,
-            "target_language": "",
-            "vac": True,
-            "vac_onnx": False,
-            "vac_chunk_size": 0.04,
-            "log_level": "DEBUG",
-            "ssl_certfile": None,
-            "ssl_keyfile": None,
-            "forwarded_allow_ips": None,
-            "transcription": True,
-            "vad": True,
-            "pcm_input": False,
-            "disable_punctuation_split" : False,
-            "diarization_backend": "sortformer",
-            "backend_policy": "simulstreaming",
-            "backend": "auto",
-        }
-        global_params = update_with_kwargs(global_params, kwargs)
+        try:
+            self._do_init(config, **kwargs)
+        except Exception:
+            # Reset singleton so a retry is possible
+            with TranscriptionEngine._lock:
+                TranscriptionEngine._instance = None
+                TranscriptionEngine._initialized = False
+            raise
 
-        transcription_common_params = {
-            "warmup_file": None,
-            "min_chunk_size": 0.1,
-            "model_size": "base",
-            "model_cache_dir": None,
-            "model_dir": None,
-            "model_path": None,
-            "lan": "auto",
-            "direct_english_translation": False,
-        }
-        transcription_common_params = update_with_kwargs(transcription_common_params, kwargs)                                            
+        with TranscriptionEngine._lock:
+            TranscriptionEngine._initialized = True
 
-        if transcription_common_params['model_size'].endswith(".en"):
-            transcription_common_params["lan"] = "en"
+    def _do_init(self, config=None, **kwargs):
+        # Handle negated kwargs from programmatic API
         if 'no_transcription' in kwargs:
-            global_params['transcription'] = not global_params['no_transcription']
+            kwargs['transcription'] = not kwargs.pop('no_transcription')
         if 'no_vad' in kwargs:
-            global_params['vad'] = not kwargs['no_vad']
+            kwargs['vad'] = not kwargs.pop('no_vad')
         if 'no_vac' in kwargs:
-            global_params['vac'] = not kwargs['no_vac']
+            kwargs['vac'] = not kwargs.pop('no_vac')
+
+        if config is None:
+            if isinstance(kwargs.get('config'), WhisperLiveKitConfig):
+                config = kwargs.pop('config')
+            else:
+                config = WhisperLiveKitConfig.from_kwargs(**kwargs)
+        self.config = config
+
+        # Backward compat: expose as self.args (Namespace-like) for AudioProcessor etc.
+        self.args = Namespace(**asdict(config))
 
-        self.args = Namespace(**{**global_params, **transcription_common_params})
-        
         self.asr = None
         self.tokenizer = None
         self.diarization = None
-        self.vac_model = None
-        
-        if self.args.vac:
-            from whisperlivekit.silero_vad_iterator import load_silero_vad
-            # Use ONNX if specified, otherwise use JIT (default)
-            use_onnx = kwargs.get('vac_onnx', False)
-            self.vac_model = load_silero_vad(onnx=use_onnx)
-        
-        backend_policy = self.args.backend_policy
-        if self.args.transcription:
-            if backend_policy == "simulstreaming":                 
+        self.vac_session = None
+
+        if config.vac:
+            from whisperlivekit.silero_vad_iterator import is_onnx_available
+
+            if is_onnx_available():
+                from whisperlivekit.silero_vad_iterator import load_onnx_session
+                self.vac_session = load_onnx_session()
+            else:
+                logger.warning(
+                    "onnxruntime not installed. VAC will use JIT model which is loaded per-session. "
+                    "For multi-user scenarios, install onnxruntime: pip install onnxruntime"
+                )
+
+        transcription_common_params = {
+            "warmup_file": config.warmup_file,
+            "min_chunk_size": config.min_chunk_size,
+            "model_size": config.model_size,
+            "model_cache_dir": config.model_cache_dir,
+            "model_dir": config.model_dir,
+            "model_path": config.model_path,
+            "lora_path": config.lora_path,
+            "lan": config.lan,
+            "direct_english_translation": config.direct_english_translation,
+        }
+
+        if config.transcription:
+            if config.backend == "voxtral-mlx":
+                from whisperlivekit.voxtral_mlx_asr import VoxtralMLXASR
+                self.tokenizer = None
+                self.asr = VoxtralMLXASR(**transcription_common_params)
+                logger.info("Using Voxtral MLX native backend")
+            elif config.backend == "voxtral":
+                from whisperlivekit.voxtral_hf_streaming import VoxtralHFStreamingASR
+                self.tokenizer = None
+                self.asr = VoxtralHFStreamingASR(**transcription_common_params)
+                logger.info("Using Voxtral HF Transformers streaming backend")
+            elif config.backend_policy == "simulstreaming":
                 simulstreaming_params = {
-                    "disable_fast_encoder": False,
-                    "custom_alignment_heads": None,
-                    "frame_threshold": 25,
-                    "beams": 1,
-                    "decoder_type": None,
-                    "audio_max_len": 20.0,
-                    "audio_min_len": 0.0,
-                    "cif_ckpt_path": None,
-                    "never_fire": False,
-                    "init_prompt": None,
-                    "static_init_prompt": None,
-                    "max_context_tokens": None,
-                    "preload_model_count": 1,
+                    "disable_fast_encoder": config.disable_fast_encoder,
+                    "custom_alignment_heads": config.custom_alignment_heads,
+                    "frame_threshold": config.frame_threshold,
+                    "beams": config.beams,
+                    "decoder_type": config.decoder_type,
+                    "audio_max_len": config.audio_max_len,
+                    "audio_min_len": config.audio_min_len,
+                    "cif_ckpt_path": config.cif_ckpt_path,
+                    "never_fire": config.never_fire,
+                    "init_prompt": config.init_prompt,
+                    "static_init_prompt": config.static_init_prompt,
+                    "max_context_tokens": config.max_context_tokens,
                 }
-                simulstreaming_params = update_with_kwargs(simulstreaming_params, kwargs)
-                
-                self.tokenizer = None        
+
+                self.tokenizer = None
                 self.asr = SimulStreamingASR(
                     **transcription_common_params,
                     **simulstreaming_params,
-                    backend=self.args.backend,
+                    backend=config.backend,
                 )
                 logger.info(
                     "Using SimulStreaming policy with %s backend",
                     getattr(self.asr, "encoder_backend", "whisper"),
                 )
             else:
-                
                 whisperstreaming_params = {
-                    "buffer_trimming": "segment",
-                    "confidence_validation": False,
-                    "buffer_trimming_sec": 15,
+                    "buffer_trimming": config.buffer_trimming,
+                    "confidence_validation": config.confidence_validation,
+                    "buffer_trimming_sec": config.buffer_trimming_sec,
                 }
-                whisperstreaming_params = update_with_kwargs(whisperstreaming_params, kwargs)
-                
+
                 self.asr = backend_factory(
-                    backend=self.args.backend,
+                    backend=config.backend,
                     **transcription_common_params,
                     **whisperstreaming_params,
                 )
@@ -133,57 +145,57 @@ class TranscriptionEngine:
                     getattr(self.asr, "backend_choice", self.asr.__class__.__name__),
                 )
 
-        if self.args.diarization:
-            if self.args.diarization_backend == "diart":
+        if config.diarization:
+            if config.diarization_backend == "diart":
                 from whisperlivekit.diarization.diart_backend import DiartDiarization
-                diart_params = {
-                    "segmentation_model": "pyannote/segmentation-3.0",
-                    "embedding_model": "pyannote/embedding",
-                }
-                diart_params = update_with_kwargs(diart_params, kwargs)
                 self.diarization_model = DiartDiarization(
-                    block_duration=self.args.min_chunk_size,
-                    **diart_params
+                    block_duration=config.min_chunk_size,
+                    segmentation_model=config.segmentation_model,
+                    embedding_model=config.embedding_model,
                 )
-            elif self.args.diarization_backend == "sortformer":
+            elif config.diarization_backend == "sortformer":
                 from whisperlivekit.diarization.sortformer_backend import SortformerDiarization
                 self.diarization_model = SortformerDiarization()
-        
+
         self.translation_model = None
-        if self.args.target_language:
-            if self.args.lan == 'auto' and backend_policy != "simulstreaming":
-                raise Exception('Translation cannot be set with language auto when transcription backend is not simulstreaming')
+        if config.target_language:
+            if config.lan == 'auto' and config.backend_policy != "simulstreaming":
+                raise ValueError('Translation cannot be set with language auto when transcription backend is not simulstreaming')
             else:
                 try:
                     from nllw import load_model
-                except:
-                    raise Exception('To use translation, you must install nllw: `pip install nllw`')
-                translation_params = { 
-                    "nllb_backend": "transformers",
-                    "nllb_size": "600M"
-                }
-                translation_params = update_with_kwargs(translation_params, kwargs)
-                self.translation_model = load_model([self.args.lan], **translation_params) #in the future we want to handle different languages for different speakers
-        TranscriptionEngine._initialized = True
+                except ImportError:
+                    raise ImportError('To use translation, you must install nllw: `pip install nllw`')
+                self.translation_model = load_model(
+                    [config.lan],
+                    nllb_backend=config.nllb_backend,
+                    nllb_size=config.nllb_size,
+                )
 
 
 def online_factory(args, asr):
-    if args.backend_policy == "simulstreaming":    
+    if getattr(args, 'backend', None) == "voxtral-mlx":
+        from whisperlivekit.voxtral_mlx_asr import VoxtralMLXOnlineProcessor
+        return VoxtralMLXOnlineProcessor(asr)
+    if getattr(args, 'backend', None) == "voxtral":
+        from whisperlivekit.voxtral_hf_streaming import VoxtralHFStreamingOnlineProcessor
+        return VoxtralHFStreamingOnlineProcessor(asr)
+    if args.backend_policy == "simulstreaming":
         from whisperlivekit.simul_whisper import SimulStreamingOnlineProcessor
-        online = SimulStreamingOnlineProcessor(asr)
-    else:
-        online = OnlineASRProcessor(asr)
-    return online
+        return SimulStreamingOnlineProcessor(asr)
+    return OnlineASRProcessor(asr)
   
   
 def online_diarization_factory(args, diarization_backend):
     if args.diarization_backend == "diart":
         online = diarization_backend
         # Not the best here, since several user/instances will share the same backend, but diart is not SOTA anymore and sortformer is recommended
-    
-    if args.diarization_backend == "sortformer":
-        from whisperlivekit.diarization.sortformer_backend import SortformerDiarizationOnline
+    elif args.diarization_backend == "sortformer":
+        from whisperlivekit.diarization.sortformer_backend import \
+            SortformerDiarizationOnline
         online = SortformerDiarizationOnline(shared_model=diarization_backend)
+    else:
+        raise ValueError(f"Unknown diarization backend: {args.diarization_backend}")
     return online
 
 

whisperlivekit/diarization/diart_backend.py

@@ -1,27 +1,23 @@
 import asyncio
-import re
-import threading
-import numpy as np
 import logging
+import threading
 import time
-from queue import SimpleQueue, Empty
+from queue import Empty, SimpleQueue
+from typing import Any, List, Tuple
 
+import diart.models as m
+import numpy as np
 from diart import SpeakerDiarization, SpeakerDiarizationConfig
 from diart.inference import StreamingInference
-from diart.sources import AudioSource
-from whisperlivekit.timed_objects import SpeakerSegment
-from diart.sources import MicrophoneAudioSource
-from rx.core import Observer
-from typing import Tuple, Any, List
+from diart.sources import AudioSource, MicrophoneAudioSource
 from pyannote.core import Annotation
-import diart.models as m
+from rx.core import Observer
+
+from whisperlivekit.diarization.utils import extract_number
+from whisperlivekit.timed_objects import SpeakerSegment
 
 logger = logging.getLogger(__name__)
 
-def extract_number(s: str) -> int:
-    m = re.search(r'\d+', s)
-    return int(m.group()) if m else None
-
 class DiarizationObserver(Observer):
     """Observer that logs all data emitted by the diarization pipeline and stores speaker segments."""
     
@@ -202,14 +198,14 @@ class DiartDiarization:
     def insert_silence(self, silence_duration):
         self.observer.global_time_offset += silence_duration
 
-    async def diarize(self, pcm_array: np.ndarray):
-        """
-        Process audio data for diarization.
-        Only used when working with WebSocketAudioSource.
-        """
+    def insert_audio_chunk(self, pcm_array: np.ndarray):
+        """Buffer audio for the next diarization step."""
         if self.custom_source:
-            self.custom_source.push_audio(pcm_array)            
-        # self.observer.clear_old_segments()        
+            self.custom_source.push_audio(pcm_array)
+
+    async def diarize(self):
+        """Return the current speaker segments from the diarization pipeline."""
+        return self.observer.get_segments()        
 
     def close(self):
         """Close the audio source."""

whisperlivekit/diarization/sortformer_backend.py

@@ -1,11 +1,12 @@
-import numpy as np
-import torch
 import logging
 import threading
 import time
 import wave
+from queue import Empty, SimpleQueue
 from typing import List, Optional
-from queue import SimpleQueue, Empty
+
+import numpy as np
+import torch
 
 from whisperlivekit.timed_objects import SpeakerSegment
 
@@ -286,15 +287,12 @@ class SortformerDiarizationOnline:
             logger.info(f"Saved {len(concatenated_audio)} samples to diarization_audio.wav")
 
 
-def extract_number(s: str) -> int:
-    """Extract number from speaker string (compatibility function)."""
-    import re
-    m = re.search(r'\d+', s)
-    return int(m.group()) if m else 0
+from whisperlivekit.diarization.utils import extract_number
 
 
 if __name__ == '__main__':
     import asyncio
+
     import librosa
     
     async def main():

whisperlivekit/diarization/utils.py

@@ -0,0 +1,7 @@
+import re
+
+
+def extract_number(s: str) -> int:
+    """Extract the first integer from a string, e.g. 'speaker_2' -> 2."""
+    m = re.search(r'\d+', s)
+    return int(m.group()) if m else 0

whisperlivekit/ffmpeg_manager.py

@@ -1,8 +1,8 @@
 import asyncio
+import contextlib
 import logging
 from enum import Enum
-from typing import Optional, Callable
-import contextlib
+from typing import Callable, Optional
 
 logger = logging.getLogger(__name__)
 logging.basicConfig(level=logging.INFO)

whisperlivekit/local_agreement/backends.py

@@ -1,34 +1,31 @@
-import sys
-import logging
 import io
-import soundfile as sf
+import logging
 import math
+import sys
 from typing import List
+
 import numpy as np
+import soundfile as sf
+
+from whisperlivekit.model_paths import detect_model_format, resolve_model_path
 from whisperlivekit.timed_objects import ASRToken
-from whisperlivekit.model_paths import resolve_model_path, model_path_and_type
 from whisperlivekit.whisper.transcribe import transcribe as whisper_transcribe
+
 logger = logging.getLogger(__name__)
 class ASRBase:
     sep = " "  # join transcribe words with this character (" " for whisper_timestamped,
               # "" for faster-whisper because it emits the spaces when needed)
 
-    def __init__(self, lan, model_size=None, cache_dir=None, model_dir=None, logfile=sys.stderr):
+    def __init__(self, lan, model_size=None, cache_dir=None, model_dir=None, lora_path=None, logfile=sys.stderr):
         self.logfile = logfile
         self.transcribe_kargs = {}
+        self.lora_path = lora_path
         if lan == "auto":
             self.original_language = None
         else:
             self.original_language = lan
         self.model = self.load_model(model_size, cache_dir, model_dir)
 
-    def with_offset(self, offset: float) -> ASRToken:
-        # This method is kept for compatibility (typically you will use ASRToken.with_offset)
-        return ASRToken(self.start + offset, self.end + offset, self.text)
-
-    def __repr__(self):
-        return f"ASRToken(start={self.start:.2f}, end={self.end:.2f}, text={self.text!r})"
-
     def load_model(self, model_size, cache_dir, model_dir):
         raise NotImplementedError("must be implemented in the child class")
 
@@ -44,24 +41,23 @@ class WhisperASR(ASRBase):
     sep = " "
 
     def load_model(self, model_size=None, cache_dir=None, model_dir=None):
-        from whisperlivekit.whisper import load_model as load_model
+        from whisperlivekit.whisper import load_model as load_whisper_model
 
         if model_dir is not None:
-            resolved_path = resolve_model_path(model_dir)
+            resolved_path = resolve_model_path(model_dir)            
             if resolved_path.is_dir():
-                pytorch_path, _, _ = model_path_and_type(resolved_path)
-                if pytorch_path is None:
+                model_info = detect_model_format(resolved_path)
+                if not model_info.has_pytorch:
                     raise FileNotFoundError(
                         f"No supported PyTorch checkpoint found under {resolved_path}"
-                    )
-                resolved_path = pytorch_path
+                    )            
             logger.debug(f"Loading Whisper model from custom path {resolved_path}")
-            return load_model(str(resolved_path))
+            return load_whisper_model(str(resolved_path), lora_path=self.lora_path)
 
         if model_size is None:
             raise ValueError("Either model_size or model_dir must be set for WhisperASR")
 
-        return load_model(model_size, download_root=cache_dir)
+        return load_whisper_model(model_size, download_root=cache_dir, lora_path=self.lora_path)
 
     def transcribe(self, audio, init_prompt=""):
         options = dict(self.transcribe_kargs)
@@ -165,8 +161,8 @@ class MLXWhisper(ASRBase):
     sep = ""
 
     def load_model(self, model_size=None, cache_dir=None, model_dir=None):
-        from mlx_whisper.transcribe import ModelHolder, transcribe
         import mlx.core as mx
+        from mlx_whisper.transcribe import ModelHolder, transcribe
 
         if model_dir is not None:
             resolved_path = resolve_model_path(model_dir)
@@ -184,22 +180,8 @@ class MLXWhisper(ASRBase):
         return transcribe
 
     def translate_model_name(self, model_name):
-        model_mapping = {
-            "tiny.en": "mlx-community/whisper-tiny.en-mlx",
-            "tiny": "mlx-community/whisper-tiny-mlx",
-            "base.en": "mlx-community/whisper-base.en-mlx",
-            "base": "mlx-community/whisper-base-mlx",
-            "small.en": "mlx-community/whisper-small.en-mlx",
-            "small": "mlx-community/whisper-small-mlx",
-            "medium.en": "mlx-community/whisper-medium.en-mlx",
-            "medium": "mlx-community/whisper-medium-mlx",
-            "large-v1": "mlx-community/whisper-large-v1-mlx",
-            "large-v2": "mlx-community/whisper-large-v2-mlx",
-            "large-v3": "mlx-community/whisper-large-v3-mlx",
-            "large-v3-turbo": "mlx-community/whisper-large-v3-turbo",
-            "large": "mlx-community/whisper-large-mlx",
-        }
-        mlx_model_path = model_mapping.get(model_name)
+        from whisperlivekit.model_mapping import MLX_MODEL_MAPPING
+        mlx_model_path = MLX_MODEL_MAPPING.get(model_name)
         if mlx_model_path:
             return mlx_model_path
         else:
@@ -224,7 +206,6 @@ class MLXWhisper(ASRBase):
             if segment.get("no_speech_prob", 0) > 0.9:
                 continue
             for word in segment.get("words", []):
-                probability=word["probability"]
                 token = ASRToken(word["start"], word["end"], word["word"])
                 tokens.append(token)
         return tokens
@@ -235,6 +216,7 @@ class MLXWhisper(ASRBase):
     def use_vad(self):
         self.transcribe_kargs["vad_filter"] = True
 
+
 class OpenaiApiASR(ASRBase):
     """Uses OpenAI's Whisper API for transcription."""
     def __init__(self, lan=None, temperature=0, logfile=sys.stderr):
@@ -246,6 +228,7 @@ class OpenaiApiASR(ASRBase):
         self.load_model()
         self.use_vad_opt = False
         self.direct_english_translation = False
+        self.task = "transcribe"
 
     def load_model(self, *args, **kwargs):
         from openai import OpenAI
@@ -291,7 +274,8 @@ class OpenaiApiASR(ASRBase):
             params["language"] = self.original_language
         if prompt:
             params["prompt"] = prompt
-        proc = self.client.audio.translations if self.task == "translate" else self.client.audio.transcriptions
+        task = self.transcribe_kargs.get("task", self.task)
+        proc = self.client.audio.translations if task == "translate" else self.client.audio.transcriptions
         transcript = proc.create(**params)
         logger.debug(f"OpenAI API processed accumulated {self.transcribed_seconds} seconds")
         return transcript

whisperlivekit/local_agreement/online_asr.py

@@ -1,7 +1,9 @@
-import sys
-import numpy as np
 import logging
-from typing import List, Tuple, Optional
+import sys
+from typing import List, Optional, Tuple
+
+import numpy as np
+
 from whisperlivekit.timed_objects import ASRToken, Sentence, Transcript
 
 logger = logging.getLogger(__name__)
@@ -134,6 +136,11 @@ class OnlineASRProcessor:
                 f"buffer_trimming_sec is set to {self.buffer_trimming_sec}, which is very long. It may cause OOM."
             )
 
+    def new_speaker(self, change_speaker):
+        """Handle speaker change event."""
+        self.process_iter()
+        self.init(offset=change_speaker.start)
+
     def init(self, offset: Optional[float] = None):
         """Initialize or reset the processing buffers."""
         self.audio_buffer = np.array([], dtype=np.float32)

whisperlivekit/local_agreement/whisper_online.py

@@ -1,18 +1,14 @@
 #!/usr/bin/env python3
-import sys
-import numpy as np
-import librosa
-from functools import lru_cache
-import time
 import logging
 import platform
-from .backends import FasterWhisperASR, MLXWhisper, WhisperASR, OpenaiApiASR
+import time
+
+from whisperlivekit.backend_support import (faster_backend_available,
+                                            mlx_backend_available)
+from whisperlivekit.model_paths import detect_model_format, resolve_model_path
 from whisperlivekit.warmup import warmup_asr
-from whisperlivekit.model_paths import resolve_model_path, model_path_and_type
-from whisperlivekit.backend_support import (
-    mlx_backend_available,
-    faster_backend_available,
-)
+
+from .backends import FasterWhisperASR, MLXWhisper, OpenaiApiASR, WhisperASR
 
 logger = logging.getLogger(__name__)
 
@@ -76,6 +72,7 @@ def backend_factory(
             model_cache_dir,
             model_dir,
             model_path,
+            lora_path,
             direct_english_translation,
             buffer_trimming,
             buffer_trimming_sec,
@@ -86,16 +83,20 @@ def backend_factory(
     backend_choice = backend
     custom_reference = model_path or model_dir
     resolved_root = None
-    pytorch_checkpoint = None
     has_mlx_weights = False
     has_fw_weights = False
+    has_pytorch = False
 
     if custom_reference:
         resolved_root = resolve_model_path(custom_reference)
         if resolved_root.is_dir():
-            pytorch_checkpoint, has_mlx_weights, has_fw_weights = model_path_and_type(resolved_root)
+            model_info = detect_model_format(resolved_root)
+            has_mlx_weights = model_info.compatible_whisper_mlx
+            has_fw_weights = model_info.compatible_faster_whisper
+            has_pytorch = model_info.has_pytorch
         else:
-            pytorch_checkpoint = resolved_root
+            # Single file provided
+            has_pytorch = True
 
     if backend_choice == "openai-api":
         logger.debug("Using OpenAI API.")
@@ -120,8 +121,8 @@ def backend_factory(
             model_override = str(resolved_root) if resolved_root is not None else None
         else:
             asr_cls = WhisperASR
-            model_override = str(pytorch_checkpoint) if pytorch_checkpoint is not None else None
-            if custom_reference and model_override is None:
+            model_override = str(resolved_root) if resolved_root is not None else None
+            if custom_reference and not has_pytorch:
                 raise FileNotFoundError(
                     f"No PyTorch checkpoint found under {resolved_root or custom_reference}"
                 )
@@ -133,12 +134,14 @@ def backend_factory(
             lan=lan,
             cache_dir=model_cache_dir,
             model_dir=model_override,
+            lora_path=lora_path if backend_choice == "whisper" else None,
         )
         e = time.time()
         logger.info(f"done. It took {round(e-t,2)} seconds.")
 
     if direct_english_translation:
         tgt_language = "en"  # Whisper translates into English
+        asr.transcribe_kargs["task"] = "translate"
     else:
         tgt_language = lan  # Whisper transcribes in this language
 
@@ -147,9 +150,9 @@ def backend_factory(
         tokenizer = create_tokenizer(tgt_language)
     else:
         tokenizer = None
-    
+
     warmup_asr(asr, warmup_file)
-    
+
     asr.confidence_validation = confidence_validation
     asr.tokenizer = tokenizer
     asr.buffer_trimming = buffer_trimming

whisperlivekit/metrics.py

@@ -0,0 +1,156 @@
+"""Lightweight ASR evaluation metrics — no external dependencies.
+
+Provides WER (Word Error Rate) computation via word-level Levenshtein distance,
+text normalization, and word-level timestamp accuracy metrics with greedy alignment.
+"""
+
+import re
+import unicodedata
+from typing import Dict, List, Optional
+
+
+def normalize_text(text: str) -> str:
+    """Normalize text for WER comparison: lowercase, strip punctuation, collapse whitespace."""
+    text = text.lower()
+    # Normalize unicode (e.g., accented chars to composed form)
+    text = unicodedata.normalize("NFC", text)
+    # Remove punctuation (keep letters, numbers, spaces, hyphens within words)
+    text = re.sub(r"[^\w\s\-']", " ", text)
+    # Collapse whitespace
+    text = re.sub(r"\s+", " ", text).strip()
+    return text
+
+
+def compute_wer(reference: str, hypothesis: str) -> Dict:
+    """Compute Word Error Rate using word-level Levenshtein edit distance.
+
+    Args:
+        reference: Ground truth transcription.
+        hypothesis: Predicted transcription.
+
+    Returns:
+        Dict with keys: wer, substitutions, insertions, deletions, ref_words, hyp_words.
+        WER can exceed 1.0 if there are more errors than reference words.
+    """
+    ref_words = normalize_text(reference).split()
+    hyp_words = normalize_text(hypothesis).split()
+
+    n = len(ref_words)
+    m = len(hyp_words)
+
+    if n == 0:
+        return {
+            "wer": 0.0 if m == 0 else float(m),
+            "substitutions": 0,
+            "insertions": m,
+            "deletions": 0,
+            "ref_words": 0,
+            "hyp_words": m,
+        }
+
+    # DP table: dp[i][j] = (edit_distance, substitutions, insertions, deletions)
+    dp = [[(0, 0, 0, 0) for _ in range(m + 1)] for _ in range(n + 1)]
+
+    for i in range(1, n + 1):
+        dp[i][0] = (i, 0, 0, i)
+    for j in range(1, m + 1):
+        dp[0][j] = (j, 0, j, 0)
+
+    for i in range(1, n + 1):
+        for j in range(1, m + 1):
+            if ref_words[i - 1] == hyp_words[j - 1]:
+                dp[i][j] = dp[i - 1][j - 1]
+            else:
+                sub = dp[i - 1][j - 1]
+                ins = dp[i][j - 1]
+                dele = dp[i - 1][j]
+
+                sub_cost = (sub[0] + 1, sub[1] + 1, sub[2], sub[3])
+                ins_cost = (ins[0] + 1, ins[1], ins[2] + 1, ins[3])
+                del_cost = (dele[0] + 1, dele[1], dele[2], dele[3] + 1)
+
+                dp[i][j] = min(sub_cost, del_cost, ins_cost, key=lambda x: x[0])
+
+    dist, subs, ins, dels = dp[n][m]
+    return {
+        "wer": dist / n,
+        "substitutions": subs,
+        "insertions": ins,
+        "deletions": dels,
+        "ref_words": n,
+        "hyp_words": m,
+    }
+
+
+def compute_timestamp_accuracy(
+    predicted: List[Dict],
+    reference: List[Dict],
+) -> Dict:
+    """Compute timestamp accuracy by aligning predicted words to reference words.
+
+    Uses greedy left-to-right alignment on normalized text. For each matched pair,
+    computes the start-time delta (predicted - reference).
+
+    Args:
+        predicted: List of dicts with keys: word, start, end.
+        reference: List of dicts with keys: word, start, end.
+
+    Returns:
+        Dict with keys: mae_start, max_delta_start, median_delta_start,
+        n_matched, n_ref, n_pred. Returns None values if no matches found.
+    """
+    if not predicted or not reference:
+        return {
+            "mae_start": None,
+            "max_delta_start": None,
+            "median_delta_start": None,
+            "n_matched": 0,
+            "n_ref": len(reference),
+            "n_pred": len(predicted),
+        }
+
+    # Normalize words for matching
+    pred_norm = [normalize_text(p["word"]) for p in predicted]
+    ref_norm = [normalize_text(r["word"]) for r in reference]
+
+    # Greedy left-to-right alignment
+    deltas_start = []
+    ref_idx = 0
+    for p_idx, p_word in enumerate(pred_norm):
+        if not p_word:
+            continue
+        # Scan forward in reference to find a match (allow small skips)
+        search_limit = min(ref_idx + 3, len(ref_norm))
+        for r_idx in range(ref_idx, search_limit):
+            if ref_norm[r_idx] == p_word:
+                delta = predicted[p_idx]["start"] - reference[r_idx]["start"]
+                deltas_start.append(delta)
+                ref_idx = r_idx + 1
+                break
+
+    if not deltas_start:
+        return {
+            "mae_start": None,
+            "max_delta_start": None,
+            "median_delta_start": None,
+            "n_matched": 0,
+            "n_ref": len(reference),
+            "n_pred": len(predicted),
+        }
+
+    abs_deltas = [abs(d) for d in deltas_start]
+    sorted_abs = sorted(abs_deltas)
+    n = len(sorted_abs)
+    if n % 2 == 1:
+        median = sorted_abs[n // 2]
+    else:
+        median = (sorted_abs[n // 2 - 1] + sorted_abs[n // 2]) / 2
+
+    return {
+        "mae_start": sum(abs_deltas) / len(abs_deltas),
+        "max_delta_start": max(abs_deltas),
+        "median_delta_start": median,
+        "n_matched": len(deltas_start),
+        "n_ref": len(reference),
+        "n_pred": len(predicted),
+    }

whisperlivekit/metrics_collector.py

@@ -0,0 +1,84 @@
+"""Lightweight runtime metrics for AudioProcessor sessions.
+
+Zero external dependencies. Negligible overhead when not queried —
+just integer increments and list appends during normal operation.
+"""
+
+import logging
+import time
+from dataclasses import dataclass, field
+from typing import Dict, List
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class SessionMetrics:
+    """Per-session metrics collected by AudioProcessor."""
+
+    session_start: float = 0.0
+    total_audio_duration_s: float = 0.0
+    total_processing_time_s: float = 0.0
+
+    # Chunk / call counters
+    n_chunks_received: int = 0
+    n_transcription_calls: int = 0
+    n_tokens_produced: int = 0
+    n_responses_sent: int = 0
+
+    # Per-call ASR latency (seconds)
+    transcription_durations: List[float] = field(default_factory=list)
+
+    # Silence
+    n_silence_events: int = 0
+    total_silence_duration_s: float = 0.0
+
+    # --- Computed properties ---
+
+    @property
+    def rtf(self) -> float:
+        """Real-time factor: processing_time / audio_duration."""
+        if self.total_audio_duration_s <= 0:
+            return 0.0
+        return self.total_processing_time_s / self.total_audio_duration_s
+
+    @property
+    def avg_latency_ms(self) -> float:
+        """Average per-call ASR latency in milliseconds."""
+        if not self.transcription_durations:
+            return 0.0
+        return (sum(self.transcription_durations) / len(self.transcription_durations)) * 1000
+
+    @property
+    def p95_latency_ms(self) -> float:
+        """95th percentile per-call ASR latency in milliseconds."""
+        if not self.transcription_durations:
+            return 0.0
+        sorted_d = sorted(self.transcription_durations)
+        idx = int(len(sorted_d) * 0.95)
+        idx = min(idx, len(sorted_d) - 1)
+        return sorted_d[idx] * 1000
+
+    def to_dict(self) -> Dict:
+        """Serialize to a plain dict (JSON-safe)."""
+        return {
+            "session_start": self.session_start,
+            "total_audio_duration_s": round(self.total_audio_duration_s, 3),
+            "total_processing_time_s": round(self.total_processing_time_s, 3),
+            "rtf": round(self.rtf, 3),
+            "n_chunks_received": self.n_chunks_received,
+            "n_transcription_calls": self.n_transcription_calls,
+            "n_tokens_produced": self.n_tokens_produced,
+            "n_responses_sent": self.n_responses_sent,
+            "avg_latency_ms": round(self.avg_latency_ms, 2),
+            "p95_latency_ms": round(self.p95_latency_ms, 2),
+            "n_silence_events": self.n_silence_events,
+            "total_silence_duration_s": round(self.total_silence_duration_s, 3),
+        }
+
+    def log_summary(self) -> None:
+        """Emit a structured log line summarising the session."""
+        self.total_processing_time_s = sum(self.transcription_durations)
+        d = self.to_dict()
+        d["session_elapsed_s"] = round(time.time() - self.session_start, 3) if self.session_start else 0
+        logger.info(f"SESSION_METRICS {d}")

whisperlivekit/model_mapping.py

@@ -0,0 +1,17 @@
+"""Shared MLX model name mapping used by both SimulStreaming and LocalAgreement backends."""
+
+MLX_MODEL_MAPPING = {
+    "tiny.en": "mlx-community/whisper-tiny.en-mlx",
+    "tiny": "mlx-community/whisper-tiny-mlx",
+    "base.en": "mlx-community/whisper-base.en-mlx",
+    "base": "mlx-community/whisper-base-mlx",
+    "small.en": "mlx-community/whisper-small.en-mlx",
+    "small": "mlx-community/whisper-small-mlx",
+    "medium.en": "mlx-community/whisper-medium.en-mlx",
+    "medium": "mlx-community/whisper-medium-mlx",
+    "large-v1": "mlx-community/whisper-large-v1-mlx",
+    "large-v2": "mlx-community/whisper-large-v2-mlx",
+    "large-v3": "mlx-community/whisper-large-v3-mlx",
+    "large-v3-turbo": "mlx-community/whisper-large-v3-turbo",
+    "large": "mlx-community/whisper-large-mlx",
+}

whisperlivekit/model_paths.py

@@ -1,49 +1,195 @@
+import json
+import re
+from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Optional, Tuple, Union
+from typing import List, Optional, Tuple, Union
+
+
+@dataclass
+class ModelInfo:
+    """Information about detected model format and files in a directory."""    
+    path: Optional[Path] = None    
+    pytorch_files: List[Path] = field(default_factory=list)
+    compatible_whisper_mlx: bool = False
+    compatible_faster_whisper: bool = False
+    
+    @property
+    def has_pytorch(self) -> bool:
+        return len(self.pytorch_files) > 0
+    
+    @property
+    def is_sharded(self) -> bool:
+        return len(self.pytorch_files) > 1
+    
+    @property
+    def primary_pytorch_file(self) -> Optional[Path]:
+        """Return the primary PyTorch file (or first shard for sharded models)."""
+        if not self.pytorch_files:
+            return None
+        return self.pytorch_files[0]
+
+
+#regex pattern for sharded model files such as: model-00001-of-00002.safetensors or pytorch_model-00001-of-00002.bin
+SHARDED_PATTERN = re.compile(r"^(.+)-(\d{5})-of-(\d{5})\.(safetensors|bin)$")
+
+FASTER_WHISPER_MARKERS = {"model.bin", "encoder.bin", "decoder.bin"}
+MLX_WHISPER_MARKERS = {"weights.npz", "weights.safetensors"}
+CT2_INDICATOR_FILES = {"vocabulary.json", "vocabulary.txt", "shared_vocabulary.json"}
+
+
+def _is_ct2_model_bin(directory: Path, filename: str) -> bool:
+    """
+    Determine if model.bin/encoder.bin/decoder.bin is a CTranslate2 model.
+    
+    CTranslate2 models have specific companion files that distinguish them
+    from PyTorch .bin files.
+    """
+    n_indicators = 0
+    for indicator in CT2_INDICATOR_FILES: #test 1
+        if (directory / indicator).exists(): 
+            n_indicators += 1
+        
+    if n_indicators == 0:
+        return False
+
+    config_path = directory / "config.json" #test 2
+    if config_path.exists():
+        try:
+            with open(config_path, "r", encoding="utf-8") as f:
+                config = json.load(f)
+            if config.get("model_type") == "whisper": #test 2
+                return False
+        except (json.JSONDecodeError, IOError):
+            pass
+    
+    return True
+
+
+def _collect_pytorch_files(directory: Path) -> List[Path]:
+    """
+    Collect all PyTorch checkpoint files from a directory.
+    
+    Handles:
+    - Single files: model.safetensors, pytorch_model.bin, *.pt
+    - Sharded files: model-00001-of-00002.safetensors, pytorch_model-00001-of-00002.bin
+    - Index-based sharded models (reads index file to find shards)
+    
+    Returns files sorted appropriately (shards in order, or single file).
+    """
+    for index_name in ["model.safetensors.index.json", "pytorch_model.bin.index.json"]:
+        index_path = directory / index_name
+        if index_path.exists():
+            try:
+                with open(index_path, "r", encoding="utf-8") as f:
+                    index_data = json.load(f)
+                weight_map = index_data.get("weight_map", {})
+                if weight_map:
+                    shard_names = sorted(set(weight_map.values()))
+                    shards = [directory / name for name in shard_names if (directory / name).exists()]
+                    if shards:
+                        return shards
+            except (json.JSONDecodeError, IOError):
+                pass
+    
+    sharded_groups = {}
+    single_files = {}
+    
+    for file in directory.iterdir():
+        if not file.is_file():
+            continue
+        
+        filename = file.name
+        suffix = file.suffix.lower()
+        
+        if filename.startswith("adapter_"):
+            continue
+        
+        match = SHARDED_PATTERN.match(filename)
+        if match:
+            base_name, shard_idx, total_shards, ext = match.groups()
+            key = (base_name, ext, int(total_shards))
+            if key not in sharded_groups:
+                sharded_groups[key] = []
+            sharded_groups[key].append((int(shard_idx), file))
+            continue
+        
+        if filename == "model.safetensors":
+            single_files[0] = file  # Highest priority
+        elif filename == "pytorch_model.bin":
+            single_files[1] = file
+        elif suffix == ".pt":
+            single_files[2] = file
+        elif suffix == ".safetensors" and not filename.startswith("adapter"):
+            single_files[3] = file
+    
+    for (base_name, ext, total_shards), shards in sharded_groups.items():
+        if len(shards) == total_shards:
+            return [path for _, path in sorted(shards)]
+    
+    for priority in sorted(single_files.keys()):
+        return [single_files[priority]]
+    
+    return []
+
+
+def detect_model_format(model_path: Union[str, Path]) -> ModelInfo:
+    """
+    Detect the model format in a given path.
+    
+    This function analyzes a file or directory to determine:
+    - What PyTorch checkpoint files are available (including sharded models)
+    - Whether the directory contains MLX Whisper weights
+    - Whether the directory contains Faster-Whisper (CTranslate2) weights
+    
+    Args:
+        model_path: Path to a model file or directory
+        
+    Returns:
+        ModelInfo with detected format information
+    """
+    path = Path(model_path)
+    info = ModelInfo(path=path)
+    
+    if path.is_file():
+        suffix = path.suffix.lower()
+        if suffix in {".pt", ".safetensors", ".bin"}:
+            info.pytorch_files = [path]
+        return info
+    
+    if not path.is_dir():
+        return info
+    
+    for file in path.iterdir():
+        if not file.is_file():
+            continue
+        
+        filename = file.name.lower()
+        
+        if filename in MLX_WHISPER_MARKERS:
+            info.compatible_whisper_mlx = True
+        
+        if filename in FASTER_WHISPER_MARKERS:
+            if _is_ct2_model_bin(path, filename):
+                info.compatible_faster_whisper = True
+    
+    info.pytorch_files = _collect_pytorch_files(path)
+    
+    return info
 
 
 def model_path_and_type(model_path: Union[str, Path]) -> Tuple[Optional[Path], bool, bool]:
     """
     Inspect the provided path and determine which model formats are available.
-
+    
+    This is a compatibility wrapper around detect_model_format().
+    
     Returns:
-        pytorch_path: Path to a PyTorch checkpoint (if present).
+        pytorch_path: Path to a PyTorch checkpoint (first shard for sharded models, or None).
         compatible_whisper_mlx: True if MLX weights exist in this folder.
-        compatible_faster_whisper: True if Faster-Whisper (ctranslate2) weights exist.
+        compatible_faster_whisper: True if Faster-Whisper (CTranslate2) weights exist.
     """
-    path = Path(model_path)
-
-    compatible_whisper_mlx = False
-    compatible_faster_whisper = False
-    pytorch_path: Optional[Path] = None
-
-    if path.is_file() and path.suffix.lower() in [".pt", ".safetensors", ".bin"]:
-        pytorch_path = path
-        return pytorch_path, compatible_whisper_mlx, compatible_faster_whisper
-
-    if path.is_dir():
-        for file in path.iterdir():
-            if not file.is_file():
-                continue
-
-            filename = file.name.lower()
-            suffix = file.suffix.lower()
-
-            if filename in {"weights.npz", "weights.safetensors"}:
-                compatible_whisper_mlx = True
-            elif filename in {"model.bin", "encoder.bin", "decoder.bin"}:
-                compatible_faster_whisper = True
-            elif suffix in {".pt", ".safetensors"}:
-                pytorch_path = file
-            elif filename == "pytorch_model.bin":
-                pytorch_path = file
-
-        if pytorch_path is None:
-            fallback = path / "pytorch_model.bin"
-            if fallback.exists():
-                pytorch_path = fallback
-
-    return pytorch_path, compatible_whisper_mlx, compatible_faster_whisper
+    info = detect_model_format(model_path)
+    return info.primary_pytorch_file, info.compatible_whisper_mlx, info.compatible_faster_whisper
 
 
 def resolve_model_path(model_path: Union[str, Path]) -> Path:
@@ -59,7 +205,7 @@ def resolve_model_path(model_path: Union[str, Path]) -> Path:
 
     try:
         from huggingface_hub import snapshot_download
-    except ImportError as exc:  # pragma: no cover - optional dependency guard
+    except ImportError as exc:
         raise FileNotFoundError(
             f"Model path '{model_path}' does not exist locally and huggingface_hub "
             "is not installed to download it."

whisperlivekit/parse_args.py

@@ -1,6 +1,7 @@
 
 from argparse import ArgumentParser
 
+
 def parse_args():
     parser = ArgumentParser(description="Whisper FastAPI Online Server")
     parser.add_argument(
@@ -105,6 +106,13 @@ def parse_args():
         default=None,
         help="Dir where Whisper model.bin and other files are saved. This option overrides --model and --model_cache_dir parameter.",
     )
+    parser.add_argument(
+        "--lora-path",
+        type=str,
+        default=None,
+        dest="lora_path",
+        help="Path or Hugging Face repo ID for LoRA adapter weights (e.g., QuentinFuxa/whisper-base-french-lora). Only works with native Whisper backend.",
+    )
     parser.add_argument(
         "--lan",
         "--language",
@@ -139,8 +147,8 @@ def parse_args():
         "--backend",
         type=str,
         default="auto",
-        choices=["auto", "mlx-whisper", "faster-whisper", "whisper", "openai-api"],
-        help="Select the Whisper backend implementation (auto: prefer MLX on macOS, otherwise Faster-Whisper, else Whisper). Use 'openai-api' with --backend-policy localagreement to call OpenAI's API.",
+        choices=["auto", "mlx-whisper", "faster-whisper", "whisper", "openai-api", "voxtral", "voxtral-mlx"],
+        help="Select the ASR backend implementation (auto: prefer MLX on macOS, otherwise Faster-Whisper, else Whisper). Use 'voxtral' for HF Transformers Voxtral (CUDA/CPU/MPS). Use 'voxtral-mlx' for native MLX Voxtral on Apple Silicon.",
     )
     parser.add_argument(
         "--no-vac",
@@ -295,14 +303,6 @@ def parse_args():
         help="Direct path to the SimulStreaming Whisper .pt model file. Overrides --model for SimulStreaming backend.",
     )
     
-    simulstreaming_group.add_argument(
-        "--preload-model-count",
-        type=int,
-        default=1,
-        dest="preload_model_count",
-        help="Optional. Number of models to preload in memory to speed up loading (set up to the expected number of concurrent instances).",
-    )
-
     simulstreaming_group.add_argument(
         "--nllb-backend",
         type=str,
@@ -318,15 +318,12 @@ def parse_args():
     )
 
     args = parser.parse_args()
-    
     args.transcription = not args.no_transcription
-    args.vad = not args.no_vad    
+    args.vad = not args.no_vad
+    args.vac = not args.no_vac
     delattr(args, 'no_transcription')
     delattr(args, 'no_vad')
+    delattr(args, 'no_vac')
 
-    if args.backend_policy == "1":
-        args.backend_policy = "simulstreaming"
-    elif args.backend_policy == "2":
-        args.backend_policy = "localagreement"
-    
-    return args
+    from whisperlivekit.config import WhisperLiveKitConfig
+    return WhisperLiveKitConfig.from_namespace(args)

whisperlivekit/silero_vad_iterator.py

@@ -1,12 +1,22 @@
-import torch
-import numpy as np
 import warnings
 from pathlib import Path
 
+import numpy as np
+import torch
+
 """
 Code is adapted from silero-vad v6: https://github.com/snakers4/silero-vad
 """
 
+def is_onnx_available() -> bool:
+    """Check if onnxruntime is installed."""
+    try:
+        import onnxruntime
+        return True
+    except ImportError:
+        return False
+
+
 def init_jit_model(model_path: str, device=torch.device('cpu')):
     """Load a JIT model from file."""
     model = torch.jit.load(model_path, map_location=device)
@@ -14,12 +24,12 @@ def init_jit_model(model_path: str, device=torch.device('cpu')):
     return model
 
 
-class OnnxWrapper():
-    """ONNX Runtime wrapper for Silero VAD model."""
+class OnnxSession():
+    """
+    Shared ONNX session for Silero VAD model (stateless).
+    """
 
     def __init__(self, path, force_onnx_cpu=False):
-        global np
-        import numpy as np
         import onnxruntime
 
         opts = onnxruntime.SessionOptions()
@@ -31,13 +41,28 @@ class OnnxWrapper():
         else:
             self.session = onnxruntime.InferenceSession(path, sess_options=opts)
 
-        self.reset_states()
+        self.path = path
         if '16k' in path:
             warnings.warn('This model support only 16000 sampling rate!')
             self.sample_rates = [16000]
         else:
             self.sample_rates = [8000, 16000]
 
+
+class OnnxWrapper():
+    """
+    ONNX Runtime wrapper for Silero VAD model with per-instance state.
+    """
+
+    def __init__(self, session: OnnxSession, force_onnx_cpu=False):
+        self._shared_session = session
+        self.sample_rates = session.sample_rates
+        self.reset_states()
+
+    @property
+    def session(self):
+        return self._shared_session.session
+
     def _validate_input(self, x, sr: int):
         if x.dim() == 1:
             x = x.unsqueeze(0)
@@ -90,7 +115,7 @@ class OnnxWrapper():
             out, state = ort_outs
             self._state = torch.from_numpy(state)
         else:
-            raise ValueError()
+            raise ValueError(f"Unsupported sampling rate {sr}. Supported: {self.sample_rates} (with sample sizes 256 for 8000, 512 for 16000)")
 
         self._context = x[..., -context_size:]
         self._last_sr = sr
@@ -100,59 +125,63 @@ class OnnxWrapper():
         return out
 
 
-def load_silero_vad(model_path: str = None, onnx: bool = False, opset_version: int = 16):
-    """
-    Load Silero VAD model (JIT or ONNX).
-    
-    Parameters
-    ----------
-    model_path : str, optional
-        Path to model file. If None, uses default bundled model.
-    onnx : bool, default False
-        Whether to use ONNX runtime (requires onnxruntime package).
-    opset_version : int, default 16
-        ONNX opset version (15 or 16). Only used if onnx=True.
-    
-    Returns
-    -------
-    model
-        Loaded VAD model (JIT or ONNX wrapper)
-    """
+def _get_onnx_model_path(model_path: str = None, opset_version: int = 16) -> Path:
+    """Get the path to the ONNX model file."""
     available_ops = [15, 16]
-    if onnx and opset_version not in available_ops:
-        raise Exception(f'Available ONNX opset_version: {available_ops}')
+    if opset_version not in available_ops:
+        raise ValueError(f'Unsupported ONNX opset_version: {opset_version}. Available: {available_ops}')
+    
     if model_path is None:
         current_dir = Path(__file__).parent
-        data_dir = current_dir / 'vad_models'
+        data_dir = current_dir / 'silero_vad_models'
         
-        if onnx:
-            if opset_version == 16:
-                model_name = 'silero_vad.onnx'
-            else:
-                model_name = f'silero_vad_16k_op{opset_version}.onnx'
+        if opset_version == 16:
+            model_name = 'silero_vad.onnx'
         else:
-            model_name = 'silero_vad.jit'
+            model_name = f'silero_vad_16k_op{opset_version}.onnx'
         
         model_path = data_dir / model_name
         
         if not model_path.exists():
             raise FileNotFoundError(
                 f"Model file not found: {model_path}\n"
-                f"Please ensure the whisperlivekit/vad_models/ directory contains the model files."
+                f"Please ensure the whisperlivekit/silero_vad_models/ directory contains the model files."
             )
     else:
         model_path = Path(model_path)
-    if onnx:
-        try:
-            model = OnnxWrapper(str(model_path), force_onnx_cpu=True)
-        except ImportError:
-            raise ImportError(
-                "ONNX runtime not available. Install with: pip install onnxruntime\n"
-                "Or use JIT model by setting onnx=False"
+    
+    return model_path
+
+
+def load_onnx_session(model_path: str = None, opset_version: int = 16, force_onnx_cpu: bool = True) -> OnnxSession:
+    """
+    Load a shared ONNX session for Silero VAD.
+    """
+    path = _get_onnx_model_path(model_path, opset_version)
+    return OnnxSession(str(path), force_onnx_cpu=force_onnx_cpu)
+
+
+def load_jit_vad(model_path: str = None):
+    """
+    Load Silero VAD model in JIT format.
+    """
+    if model_path is None:
+        current_dir = Path(__file__).parent
+        data_dir = current_dir / 'silero_vad_models'
+        model_name = 'silero_vad.jit'
+        
+        model_path = data_dir / model_name
+        
+        if not model_path.exists():
+            raise FileNotFoundError(
+                f"Model file not found: {model_path}\n"
+                f"Please ensure the whisperlivekit/silero_vad_models/ directory contains the model files."
             )
     else:
-        model = init_jit_model(str(model_path))
+        model_path = Path(model_path)
 
+    model = init_jit_model(str(model_path))
+ 
     return model
 
 
@@ -226,8 +255,8 @@ class VADIterator:
         if not torch.is_tensor(x):
             try:
                 x = torch.Tensor(x)
-            except:
-                raise TypeError("Audio cannot be casted to tensor. Cast it manually")
+            except (ValueError, TypeError, RuntimeError) as exc:
+                raise TypeError("Audio cannot be cast to tensor. Cast it manually") from exc
 
         window_size_samples = len(x[0]) if x.dim() == 2 else len(x)
         self.current_sample += window_size_samples
@@ -276,19 +305,22 @@ class FixedVADIterator(VADIterator):
             elif r is not None:
                 if "end" in r:
                     ret["end"] = r["end"]
-                if "start" in r and "end" in ret:
-                    del ret["end"]
+                if "start" in r:
+                    ret["start"] = r["start"]
+                    if "end" in ret:
+                        del ret["end"]
         return ret if ret != {} else None
 
 
 if __name__ == "__main__":
-    model = load_silero_vad(onnx=False)
-    vad = FixedVADIterator(model)
-    
+    # vad = FixedVADIterator(load_jit_vad())
+    vad = FixedVADIterator(OnnxWrapper(session=load_onnx_session()))
+
     audio_buffer = np.array([0] * 512, dtype=np.float32)
     result = vad(audio_buffer)
     print(f"   512 samples: {result}")
     
     # test with 511 samples
     audio_buffer = np.array([0] * 511, dtype=np.float32)
-    result = vad(audio_buffer)
+    result = vad(audio_buffer)
+    print(f"   511 samples: {result}")

whisperlivekit/simul_whisper/align_att_base.py

@@ -0,0 +1,552 @@
+"""Abstract base class for AlignAtt streaming decoders (PyTorch & MLX)."""
+import logging
+from abc import ABC, abstractmethod
+from typing import Any, List, Optional, Tuple
+
+from whisperlivekit.timed_objects import ASRToken
+from whisperlivekit.whisper import DecodingOptions, tokenizer
+
+from .config import AlignAttConfig
+
+DEC_PAD = 50257
+logger = logging.getLogger(__name__)
+
+
+class AlignAttBase(ABC):
+    """
+    Abstract base class for AlignAtt streaming decoders.
+
+    Provides shared logic for both PyTorch and MLX implementations:
+    - Properties (speaker, global_time_offset)
+    - Pure-Python methods (warmup, trim_context, refresh_segment, etc.)
+    - Template infer() with abstract hooks for tensor-specific operations
+    - Post-decode logic (token splitting, timestamped word building)
+
+    Subclasses must implement ~20 abstract methods for tensor-specific ops.
+    """
+
+    # === Properties ===
+
+    @property
+    def speaker(self):
+        return self.state.speaker
+
+    @speaker.setter
+    def speaker(self, value):
+        self.state.speaker = value
+
+    @property
+    def global_time_offset(self):
+        return self.state.global_time_offset
+
+    @global_time_offset.setter
+    def global_time_offset(self, value):
+        self.state.global_time_offset = value
+
+    # === Constructor helpers ===
+
+    def _base_init(self, cfg: AlignAttConfig, model):
+        """Common initialization — call from subclass __init__."""
+        self.model = model
+        self.cfg = cfg
+        self.decode_options = DecodingOptions(
+            language=cfg.language,
+            without_timestamps=True,
+            task=cfg.task,
+        )
+        self.tokenizer_is_multilingual = cfg.tokenizer_is_multilingual
+        self.max_text_len = model.dims.n_text_ctx
+        self.num_decoder_layers = len(model.decoder.blocks)
+        if cfg.max_context_tokens is None:
+            self.max_context_tokens = self.max_text_len
+        else:
+            self.max_context_tokens = cfg.max_context_tokens
+
+    def _init_state_common(self, cfg: AlignAttConfig):
+        """Common state initialization — call from subclass _init_state."""
+        self.create_tokenizer(cfg.language if cfg.language != "auto" else None)
+        self.state.tokenizer = self.tokenizer
+        self.state.detected_language = cfg.language if cfg.language != "auto" else None
+        self.state.global_time_offset = 0.0
+        self.state.last_attend_frame = -cfg.rewind_threshold
+        self.state.speaker = -1
+
+    # === Shared concrete methods ===
+
+    def warmup(self, audio):
+        try:
+            self.insert_audio(audio)
+            self.infer(is_last=True)
+            self.refresh_segment(complete=True)
+            logger.info("Model warmed up successfully")
+        except Exception as e:
+            logger.exception(f"Model warmup failed: {e}")
+
+    def create_tokenizer(self, language=None):
+        self.tokenizer = tokenizer.get_tokenizer(
+            multilingual=self.tokenizer_is_multilingual,
+            language=language,
+            num_languages=self.model.num_languages,
+            task=self.decode_options.task,
+        )
+        self.state.tokenizer = self.tokenizer
+
+    def trim_context(self):
+        logger.info("Trimming context")
+        c = len(self.state.context.as_token_ids()) - len(self.state.context.prefix_token_ids)
+        logger.info(f"Context text: {self.state.context.as_text()}")
+        l = sum(t.shape[1] for t in self.state.tokens) + c
+        after = 0 if self.cfg.static_init_prompt is None else len(self.cfg.static_init_prompt)
+        while c > self.max_context_tokens or l > self.max_text_len - 20:
+            t = self.state.context.trim_words(after=after)
+            l -= t
+            c -= t
+            logger.debug(f"len {l}, c {c}, max_context_tokens {self.max_context_tokens}")
+            if t == 0:
+                break
+        logger.info(f"Context after trim: {self.state.context.text} (len: {l})")
+
+    def refresh_segment(self, complete=False):
+        logger.debug("Refreshing segment:")
+        self.init_tokens()
+        self.state.last_attend_frame = -self.cfg.rewind_threshold
+        self.state.cumulative_time_offset = 0.0
+        self.init_context()
+        logger.debug(f"Context: {self.state.context}")
+        if not complete and len(self.state.segments) > 2:
+            self.state.segments = self.state.segments[-2:]
+        else:
+            logger.debug("removing all segments.")
+            self.state.segments = []
+        self.state.log_segments += 1
+        self.state.pending_incomplete_tokens = []
+        self.state.pending_retries = 0
+
+    def segments_len(self):
+        return sum(s.shape[0] for s in self.state.segments) / 16000
+
+    def _apply_minseglen(self):
+        segments_len = self.segments_len()
+        if segments_len < self.cfg.audio_min_len:
+            logger.debug("waiting for next segment")
+            return False
+        return True
+
+    def _clean_cache(self):
+        self.state.clean_cache()
+
+    def debug_print_tokens(self, tokens):
+        for i in range(min(self.cfg.beam_size, tokens.shape[0])):
+            logger.debug(self.tokenizer.decode_with_timestamps(tokens[i].tolist()))
+
+    # === Language detection ===
+
+    def _detect_language_if_needed(self, encoder_feature):
+        if (
+            self.cfg.language == "auto"
+            and self.state.detected_language is None
+            and self.state.first_timestamp
+        ):
+            seconds_since_start = self.segments_len() - self.state.first_timestamp
+            if seconds_since_start >= 2.0:
+                language_tokens, language_probs = self.lang_id(encoder_feature)
+                top_lan, p = max(language_probs[0].items(), key=lambda x: x[1])
+                print(f"Detected language: {top_lan} with p={p:.4f}")
+                self.create_tokenizer(top_lan)
+                self.state.last_attend_frame = -self.cfg.rewind_threshold
+                self.state.cumulative_time_offset = 0.0
+                self.init_tokens()
+                self.init_context()
+                self.state.detected_language = top_lan
+                logger.info(f"Tokenizer language: {self.tokenizer.language}")
+
+    # === Template infer() ===
+
+    def infer(self, is_last=False):
+        """Main inference — template method calling abstract hooks for tensor ops."""
+        new_segment = True
+
+        if len(self.state.segments) == 0:
+            logger.debug("No segments, nothing to do")
+            return []
+        if not self._apply_minseglen():
+            logger.debug(f"applied minseglen {self.cfg.audio_min_len} > {self.segments_len()}.")
+            return []
+
+        input_segments = self._concat_segments()
+        encoder_feature, content_mel_len = self._encode(input_segments)
+        self._evaluate(encoder_feature)
+
+        self._detect_language_if_needed(encoder_feature)
+        self.trim_context()
+        current_tokens = self._current_tokens()
+
+        fire_detected = self.fire_at_boundary(encoder_feature[:, :content_mel_len, :])
+
+        sum_logprobs = self._init_sum_logprobs()
+        completed = False
+        token_len_before = current_tokens.shape[1]
+        l_absolute_timestamps = []
+        accumulated_cross_attns = []
+
+        audio_duration_s = self.segments_len()
+        max_tokens = max(50, int(audio_duration_s * 15 * 1.5))
+        tokens_produced = 0
+        most_attended_frame = None
+
+        while not completed and current_tokens.shape[1] < self.max_text_len:
+            tokens_produced += 1
+            if tokens_produced > max_tokens:
+                logger.warning(
+                    f"[Loop Detection] Too many tokens ({tokens_produced}) "
+                    f"for {audio_duration_s:.2f}s audio. Breaking."
+                )
+                current_tokens = current_tokens[:, :token_len_before]
+                break
+
+            tokens_for_logits = current_tokens if new_segment else current_tokens[:, -1:]
+            logits, cross_attns = self._get_logits_and_cross_attn(
+                tokens_for_logits, encoder_feature
+            )
+            self._evaluate(logits)
+
+            accumulated_cross_attns.append(cross_attns)
+            if len(accumulated_cross_attns) > 16:
+                accumulated_cross_attns = accumulated_cross_attns[-16:]
+
+            if new_segment and self._check_no_speech(logits):
+                break
+
+            logits = logits[:, -1, :]
+
+            if new_segment:
+                logits = self._suppress_blank_tokens(logits)
+            new_segment = False
+
+            logits = self._apply_token_suppression(logits)
+            logits = self._apply_dry_penalty(logits, current_tokens)
+            current_tokens, completed = self._update_tokens(
+                current_tokens, logits, sum_logprobs
+            )
+            self._evaluate(current_tokens)
+
+            logger.debug(f"Decoding completed: {completed}")
+            self.debug_print_tokens(current_tokens)
+
+            attn = self._process_cross_attention(accumulated_cross_attns, content_mel_len)
+            frames_list, most_attended_frame = self._get_attended_frames(attn)
+
+            absolute_timestamps = [
+                (frame * 0.02 + self.state.cumulative_time_offset)
+                for frame in frames_list
+            ]
+            l_absolute_timestamps.append(absolute_timestamps[0])
+            logger.debug(f"Absolute timestamps: {absolute_timestamps}")
+
+            if completed:
+                current_tokens = current_tokens[:, :-1]
+                break
+
+            # Rewind check
+            if (
+                not is_last
+                and self.state.last_attend_frame - most_attended_frame
+                > self.cfg.rewind_threshold
+            ):
+                if current_tokens.shape[1] > 1 and self._is_special_token(current_tokens):
+                    logger.debug("omit rewinding from special tokens")
+                    self.state.last_attend_frame = most_attended_frame
+                else:
+                    logger.debug(
+                        f"[rewind detected] current: {most_attended_frame}, "
+                        f"last: {self.state.last_attend_frame}"
+                    )
+                    self.state.last_attend_frame = -self.cfg.rewind_threshold
+                    current_tokens = self._rewind_tokens()
+                    break
+            else:
+                self.state.last_attend_frame = most_attended_frame
+
+            if content_mel_len - most_attended_frame <= (
+                4 if is_last else self.cfg.frame_threshold
+            ):
+                logger.debug(
+                    f"attention reaches the end: {most_attended_frame}/{content_mel_len}"
+                )
+                current_tokens = current_tokens[:, :-1]
+                break
+
+        # Post-decode: split tokens and build timestamped words
+        tokens_to_split = self._tokens_to_list(current_tokens, token_len_before)
+        if self.state.pending_incomplete_tokens:
+            logger.debug(
+                f"[UTF-8 Fix] Prepending {len(self.state.pending_incomplete_tokens)} "
+                f"pending tokens: {self.state.pending_incomplete_tokens}"
+            )
+            tokens_to_split = self.state.pending_incomplete_tokens + tokens_to_split
+
+        new_hypothesis, split_words, split_tokens = self._split_tokens(
+            tokens_to_split, fire_detected, is_last
+        )
+
+        new_tokens_tensor = self._make_new_tokens_tensor(new_hypothesis)
+        self.state.tokens.append(new_tokens_tensor)
+        logger.info(f"Output: {self.tokenizer.decode(new_hypothesis)}")
+
+        self._clean_cache()
+
+        if len(l_absolute_timestamps) >= 2 and self.state.first_timestamp is None:
+            self.state.first_timestamp = l_absolute_timestamps[0]
+
+        timestamped_words = self._build_timestamped_words(
+            split_words, split_tokens, l_absolute_timestamps
+        )
+        self._handle_pending_tokens(split_words, split_tokens)
+
+        return timestamped_words
+
+    # === Post-decode shared helpers ===
+
+    def _split_tokens(self, tokens_list, fire_detected, is_last):
+        """Split token list into words. Returns (hypothesis, split_words, split_tokens)."""
+        if fire_detected or is_last:
+            new_hypothesis = tokens_list
+            split_words, split_tokens = self.tokenizer.split_to_word_tokens(new_hypothesis)
+        else:
+            split_words, split_tokens = self.tokenizer.split_to_word_tokens(tokens_list)
+            if len(split_words) > 1:
+                new_hypothesis = [i for sublist in split_tokens[:-1] for i in sublist]
+            else:
+                new_hypothesis = []
+        return new_hypothesis, split_words, split_tokens
+
+    def _build_timestamped_words(self, split_words, split_tokens, l_absolute_timestamps):
+        """Build list of timestamped ASRToken from split words."""
+        timestamped_words = []
+        timestamp_idx = 0
+        replacement_char = "\ufffd"
+
+        for word, word_tokens in zip(split_words, split_tokens):
+            if replacement_char in word:
+                cleaned = word.replace(replacement_char, "")
+                if not cleaned.strip():
+                    logger.debug(f"[UTF-8 Filter] Skipping: {repr(word)}")
+                    timestamp_idx += len(word_tokens)
+                    continue
+                logger.debug(f"[UTF-8 Filter] Cleaned {repr(word)} -> {repr(cleaned)}")
+                word = cleaned
+
+            try:
+                current_timestamp = l_absolute_timestamps[timestamp_idx]
+            except IndexError:
+                logger.warning(
+                    f"Timestamp index {timestamp_idx} out of range, using last timestamp"
+                )
+                current_timestamp = (
+                    l_absolute_timestamps[-1] if l_absolute_timestamps else 0.0
+                )
+            timestamp_idx += len(word_tokens)
+
+            timestamp_entry = ASRToken(
+                start=round(current_timestamp, 2),
+                end=round(current_timestamp + 0.1, 2),
+                text=word,
+                speaker=self.state.speaker,
+                detected_language=self.state.detected_language,
+            ).with_offset(self.state.global_time_offset)
+            timestamped_words.append(timestamp_entry)
+
+        return timestamped_words
+
+    def _handle_pending_tokens(self, split_words, split_tokens):
+        """Handle incomplete UTF-8 tokens for next chunk."""
+        MAX_PENDING_TOKENS = 10
+        MAX_PENDING_RETRIES = 2
+        replacement_char = "\ufffd"
+
+        if split_words and replacement_char in split_words[-1]:
+            self.state.pending_retries += 1
+            if self.state.pending_retries > MAX_PENDING_RETRIES:
+                logger.warning(
+                    f"[UTF-8 Fix] Dropping {len(split_tokens[-1])} incomplete tokens "
+                    f"after {MAX_PENDING_RETRIES} retries (won't resolve)"
+                )
+                self.state.pending_incomplete_tokens = []
+                self.state.pending_retries = 0
+            elif len(split_tokens[-1]) <= MAX_PENDING_TOKENS:
+                self.state.pending_incomplete_tokens = split_tokens[-1]
+                logger.debug(
+                    f"[UTF-8 Fix] Holding {len(self.state.pending_incomplete_tokens)} "
+                    f"incomplete tokens for next chunk (retry {self.state.pending_retries})"
+                )
+            else:
+                logger.warning(
+                    f"[UTF-8 Fix] Skipping {len(split_tokens[-1])} tokens "
+                    f"(exceeds limit of {MAX_PENDING_TOKENS}, likely hallucination)"
+                )
+                self.state.pending_incomplete_tokens = []
+                self.state.pending_retries = 0
+        else:
+            self.state.pending_incomplete_tokens = []
+            self.state.pending_retries = 0
+
+    # === Repetition penalty ===
+
+    def _apply_dry_penalty(self, logits, current_tokens):
+        """DRY penalty v0: penalize tokens that would extend a verbatim repetition.
+        See https://github.com/oobabooga/text-generation-webui/pull/5677
+
+        Scans the decoded sequence for positions where the current suffix already
+        appeared --> for each such match, the token that followed it in the past is
+        penalised exponentially with the match length
+        """
+        eot = self.tokenizer.eot
+        seq = current_tokens[0].tolist()
+        if len(seq) < 5:
+            return logits
+
+        last = seq[-1]
+        if last >= eot:
+            return logits
+
+        penalties = {}
+        for i in range(len(seq) - 2, -1, -1):
+            if seq[i] != last:
+                continue
+            next_tok = seq[i + 1]
+            if next_tok >= eot:
+                continue
+
+            length = 1
+            while length < 50:
+                j, k = i - length, len(seq) - 1 - length
+                if j < 0 or k <= i:
+                    break
+                if seq[j] != seq[k] or seq[j] >= eot:
+                    break
+                length += 1
+
+            if next_tok not in penalties or length > penalties[next_tok]:
+                penalties[next_tok] = length
+
+        if penalties:
+            max_len = max(penalties.values())
+            if max_len >= 4:
+                logger.debug(f"[DRY] penalising {len(penalties)} tokens (longest match: {max_len})")
+            for tok, length in penalties.items():
+                if length >= 2:
+                    logits[:, tok] = logits[:, tok] - 1.0 * 2.0 ** (length - 2)
+
+        return logits
+
+    # === Abstract methods — subclass must implement ===
+
+    @abstractmethod
+    def _init_state(self, cfg: AlignAttConfig):
+        """Initialize per-session decoder state."""
+        ...
+
+    @abstractmethod
+    def init_tokens(self):
+        """Initialize token sequence with framework-specific tensors."""
+        ...
+
+    @abstractmethod
+    def init_context(self):
+        """Initialize context buffer with framework-specific TokenBuffer."""
+        ...
+
+    @abstractmethod
+    def insert_audio(self, segment=None):
+        """Insert audio segment into buffer."""
+        ...
+
+    @abstractmethod
+    def _current_tokens(self):
+        """Build current token tensor for decoding."""
+        ...
+
+    @abstractmethod
+    def fire_at_boundary(self, feature):
+        """Check if we should fire at word boundary."""
+        ...
+
+    @abstractmethod
+    def lang_id(self, encoder_features):
+        """Language detection from encoder features. Returns (tokens, probs)."""
+        ...
+
+    @abstractmethod
+    def _concat_segments(self):
+        """Concatenate audio segments into single array/tensor."""
+        ...
+
+    @abstractmethod
+    def _encode(self, input_segments):
+        """Encode audio. Returns (encoder_feature, content_mel_len)."""
+        ...
+
+    @abstractmethod
+    def _init_sum_logprobs(self):
+        """Create zero sum_logprobs tensor for beam search."""
+        ...
+
+    @abstractmethod
+    def _get_logits_and_cross_attn(self, tokens, encoder_feature):
+        """Get logits and cross-attention from decoder. Returns (logits, cross_attns)."""
+        ...
+
+    @abstractmethod
+    def _check_no_speech(self, logits):
+        """Check no_speech probability at start of segment. Returns True to break."""
+        ...
+
+    @abstractmethod
+    def _suppress_blank_tokens(self, logits):
+        """Suppress blank/EOT tokens at segment start. Returns modified logits."""
+        ...
+
+    @abstractmethod
+    def _apply_token_suppression(self, logits):
+        """Apply general token suppression. Returns modified logits."""
+        ...
+
+    @abstractmethod
+    def _update_tokens(self, current_tokens, logits, sum_logprobs):
+        """Update tokens via decoder. Returns (current_tokens, completed)."""
+        ...
+
+    @abstractmethod
+    def _process_cross_attention(self, accumulated_cross_attns, content_mel_len):
+        """Process cross-attention for alignment. Returns attention tensor."""
+        ...
+
+    @abstractmethod
+    def _get_attended_frames(self, attn):
+        """Get most attended frames. Returns (frames_as_python_list, first_frame_int)."""
+        ...
+
+    @abstractmethod
+    def _is_special_token(self, current_tokens):
+        """Check if second-to-last token is a special token (>= DEC_PAD)."""
+        ...
+
+    @abstractmethod
+    def _rewind_tokens(self):
+        """Concatenate state tokens for rewind. Returns token tensor."""
+        ...
+
+    @abstractmethod
+    def _tokens_to_list(self, current_tokens, start_col):
+        """Extract tokens as Python list from start_col onwards."""
+        ...
+
+    @abstractmethod
+    def _make_new_tokens_tensor(self, hypothesis):
+        """Create tensor from hypothesis token list, repeated for beam search."""
+        ...
+
+    @abstractmethod
+    def _evaluate(self, tensor):
+        """Evaluate lazy tensor (mx.eval for MLX, no-op for PyTorch)."""
+        ...

whisperlivekit/simul_whisper/backend.py

@@ -1,33 +1,34 @@
-import sys
-import numpy as np
+import gc
 import logging
-from typing import List, Tuple, Optional
+import os
 import platform
-from whisperlivekit.timed_objects import ASRToken, Transcript, ChangeSpeaker
+import sys
+from pathlib import Path
+from typing import List, Optional, Tuple
+
+import numpy as np
+import torch
+
+from whisperlivekit.backend_support import (faster_backend_available,
+                                            mlx_backend_available)
+from whisperlivekit.model_paths import detect_model_format, resolve_model_path
+from whisperlivekit.simul_whisper.config import AlignAttConfig
+from whisperlivekit.simul_whisper.simul_whisper import AlignAtt
+from whisperlivekit.timed_objects import ASRToken, ChangeSpeaker, Transcript
 from whisperlivekit.warmup import load_file
 from whisperlivekit.whisper import load_model, tokenizer
 from whisperlivekit.whisper.audio import TOKENS_PER_SECOND
-import os
-import gc
-from pathlib import Path
-from whisperlivekit.model_paths import model_path_and_type, resolve_model_path
-from whisperlivekit.backend_support import (
-    mlx_backend_available,
-    faster_backend_available,
-)
-
-import torch
-from whisperlivekit.simul_whisper.config import AlignAttConfig
-from whisperlivekit.simul_whisper.simul_whisper import AlignAtt
 
 logger = logging.getLogger(__name__)
 
 
 HAS_MLX_WHISPER = mlx_backend_available(warn_on_missing=True)
 if HAS_MLX_WHISPER:
-    from .mlx_encoder import mlx_model_mapping, load_mlx_encoder
+    from .mlx_encoder import load_mlx_encoder, load_mlx_model, mlx_model_mapping
+    from .mlx import MLXAlignAtt
 else:
     mlx_model_mapping = {}
+    MLXAlignAtt = None
 HAS_FASTER_WHISPER = faster_backend_available(warn_on_missing=not HAS_MLX_WHISPER)
 if HAS_FASTER_WHISPER:
     from faster_whisper import WhisperModel
@@ -37,32 +38,30 @@ else:
 MIN_DURATION_REAL_SILENCE = 5
 
 class SimulStreamingOnlineProcessor:
+    """Online processor for SimulStreaming ASR."""
     SAMPLING_RATE = 16000
 
-    def __init__(
-        self,
-        asr,
-        logfile=sys.stderr,
-    ):        
+    def __init__(self, asr, logfile=sys.stderr):
         self.asr = asr
         self.logfile = logfile
         self.end = 0.0
         self.buffer = []
-        self.committed: List[ASRToken] = []
-        self.last_result_tokens: List[ASRToken] = []
-        self.load_new_backend()
+        self.model = self._create_alignatt()
         
-        #can be moved
         if asr.tokenizer:
             self.model.tokenizer = asr.tokenizer
+            self.model.state.tokenizer = asr.tokenizer
 
-    def load_new_backend(self):
-        model = self.asr.get_new_model_instance()
-        self.model = AlignAtt(
-            cfg=self.asr.cfg,
-            loaded_model=model,
-            mlx_encoder=self.asr.mlx_encoder,
-            fw_encoder=self.asr.fw_encoder,
+    def _create_alignatt(self):
+        """Create the AlignAtt decoder instance based on ASR mode."""
+        if self.asr.use_full_mlx and HAS_MLX_WHISPER:
+            return MLXAlignAtt(cfg=self.asr.cfg, mlx_model=self.asr.mlx_model)
+        else:
+            return AlignAtt(
+                cfg=self.asr.cfg,
+                loaded_model=self.asr.shared_model,
+                mlx_encoder=self.asr.mlx_encoder,
+                fw_encoder=self.asr.fw_encoder,
             )
 
     def start_silence(self):
@@ -70,35 +69,36 @@ class SimulStreamingOnlineProcessor:
         return tokens, processed_upto
 
     def end_silence(self, silence_duration, offset):
-        """
-        If silences are > MIN_DURATION_REAL_SILENCE, we do a complete context clear. Otherwise, we just insert a small silence and shift the last_attend_frame
-        """
+        """Handle silence period."""
         self.end += silence_duration
         long_silence = silence_duration >= MIN_DURATION_REAL_SILENCE
         if not long_silence:
             gap_len = int(16000 * silence_duration)
             if gap_len > 0:
-                gap_silence = torch.zeros(gap_len)
+                if self.asr.use_full_mlx:
+                    gap_silence = np.zeros(gap_len, dtype=np.float32)
+                else:
+                    gap_silence = torch.zeros(gap_len)
                 self.model.insert_audio(gap_silence)
         if long_silence:
             self.model.refresh_segment(complete=True)
             self.model.global_time_offset = silence_duration + offset
 
-
-        
     def insert_audio_chunk(self, audio: np.ndarray, audio_stream_end_time):
         """Append an audio chunk to be processed by SimulStreaming."""
-            
-        # Convert numpy array to torch tensor
-        audio_tensor = torch.from_numpy(audio).float()
-        self.end = audio_stream_end_time #Only to be aligned with what happens in whisperstreaming backend.
-        self.model.insert_audio(audio_tensor)
+        self.end = audio_stream_end_time
+        if self.asr.use_full_mlx:
+            self.model.insert_audio(audio)
+        else:
+            audio_tensor = torch.from_numpy(audio).float()
+            self.model.insert_audio(audio_tensor)
 
     def new_speaker(self, change_speaker: ChangeSpeaker):
-            self.process_iter(is_last=True)
-            self.model.refresh_segment(complete=True)
-            self.model.speaker = change_speaker.speaker
-            self.global_time_offset = change_speaker.start
+        """Handle speaker change event."""
+        self.process_iter(is_last=True)
+        self.model.refresh_segment(complete=True)
+        self.model.speaker = change_speaker.speaker
+        self.model.global_time_offset = change_speaker.start
             
     def get_buffer(self):
         concat_buffer = Transcript.from_tokens(tokens= self.buffer, sep='')
@@ -112,15 +112,16 @@ class SimulStreamingOnlineProcessor:
         """
         try:
             timestamped_words = self.model.infer(is_last=is_last)
-            if self.model.cfg.language == "auto" and timestamped_words and timestamped_words[0].detected_language == None:
+            
+            if not timestamped_words:
+                return [], self.end
+            
+            if self.model.cfg.language == "auto" and timestamped_words[0].detected_language is None:
                 self.buffer.extend(timestamped_words)
                 return [], self.end
             
-            self.committed.extend(timestamped_words)
             self.buffer = []
             return timestamped_words, self.end
-
-            
         except Exception as e:
             logger.exception(f"SimulStreaming processing error: {e}")
             return [], self.end
@@ -128,6 +129,10 @@ class SimulStreamingOnlineProcessor:
     def warmup(self, audio, init_prompt=""):
         """Warmup the SimulStreaming model."""
         try:
+            if self.asr.use_full_mlx:
+                # MLX mode: ensure numpy array
+                if hasattr(audio, 'numpy'):
+                    audio = audio.numpy()
             self.model.insert_audio(audio)
             self.model.infer(True)
             self.model.refresh_segment(complete=True)
@@ -136,14 +141,15 @@ class SimulStreamingOnlineProcessor:
             logger.exception(f"SimulStreaming warmup failed: {e}")
 
     def __del__(self):
-        # free the model and add a new model to stack.
-        # del self.model
         gc.collect()
-        torch.cuda.empty_cache()
-        # self.asr.new_model_to_stack()
-        self.model.remove_hooks()
+        if not getattr(self.asr, 'use_full_mlx', True) and torch is not None:
+            try:
+                torch.cuda.empty_cache()
+            except Exception:
+                pass
 
-class SimulStreamingASR():
+
+class SimulStreamingASR:
     """SimulStreaming backend with AlignAtt policy."""
     sep = ""
 
@@ -160,35 +166,25 @@ class SimulStreamingASR():
         self.fast_encoder = False
         self._resolved_model_path = None
         self.encoder_backend = "whisper"
+        self.use_full_mlx = getattr(self, "use_full_mlx", False)
         preferred_backend = getattr(self, "backend", "auto")
-        self.pytorch_path, compatible_whisper_mlx, compatible_faster_whisper = None, True, True
+        compatible_whisper_mlx, compatible_faster_whisper = True, True
+        
         if self.model_path:
             resolved_model_path = resolve_model_path(self.model_path)
             self._resolved_model_path = resolved_model_path
             self.model_path = str(resolved_model_path)
-            self.pytorch_path, compatible_whisper_mlx, compatible_faster_whisper = model_path_and_type(resolved_model_path)
-            if self.pytorch_path:
-                self.model_name = self.pytorch_path.stem
-            else:
-                self.model_name = Path(self.model_path).stem
+            
+            model_info = detect_model_format(resolved_model_path)
+            compatible_whisper_mlx = model_info.compatible_whisper_mlx
+            compatible_faster_whisper = model_info.compatible_faster_whisper
+            
+            if not self.use_full_mlx and not model_info.has_pytorch:
                 raise FileNotFoundError(
                     f"No PyTorch checkpoint (.pt/.bin/.safetensors) found under {self.model_path}"
-                )
+                )            
+            self.model_name = resolved_model_path.name if resolved_model_path.is_dir() else resolved_model_path.stem
         elif self.model_size is not None:
-            model_mapping = {
-                'tiny': './tiny.pt',
-                'base': './base.pt',
-                'small': './small.pt',
-                'medium': './medium.pt',
-                'medium.en': './medium.en.pt',
-                'large-v1': './large-v1.pt',
-                'base.en': './base.en.pt',
-                'small.en': './small.en.pt',
-                'tiny.en': './tiny.en.pt',
-                'large-v2': './large-v2.pt',
-                'large-v3': './large-v3.pt',
-                'large': './large-v3.pt'
-            }
             self.model_name = self.model_size
         else:
             raise ValueError("Either model_size or model_path must be specified for SimulStreaming.")
@@ -203,6 +199,13 @@ class SimulStreamingASR():
         self.fast_encoder = self.encoder_backend in ("mlx-whisper", "faster-whisper")
         if self.encoder_backend == "whisper":
             self.disable_fast_encoder = True
+        
+        # MLX full decoder disabled by default — MLXAlignAtt has known issues
+        # with token generation after punctuation. Users can opt-in with
+        # --use-full-mlx if they want to test it.
+        # if self.encoder_backend == "mlx-whisper" and platform.system() == "Darwin":
+        #     if not hasattr(self, '_full_mlx_disabled'):
+        #         self.use_full_mlx = True
                     
         self.cfg = AlignAttConfig(
                 tokenizer_is_multilingual= is_multilingual,
@@ -214,7 +217,7 @@ class SimulStreamingASR():
                 cif_ckpt_path=self.cif_ckpt_path,
                 decoder_type="beam",
                 beam_size=self.beams,
-                task=self.direct_english_translation,
+                task="translate" if self.direct_english_translation else "transcribe",
                 never_fire=self.never_fire,
                 init_prompt=self.init_prompt,
                 max_context_tokens=self.max_context_tokens,
@@ -226,24 +229,37 @@ class SimulStreamingASR():
             self.tokenizer = self.set_translate_task()
         else:
             self.tokenizer = None
+
+        self.mlx_encoder, self.fw_encoder, self.mlx_model = None, None, None
+        self.shared_model = None
         
-        
-            
-    
-        self.mlx_encoder, self.fw_encoder = None, None
-        if self.encoder_backend == "mlx-whisper":
-            print('Simulstreaming will use MLX whisper to increase encoding speed.')
+        if self.use_full_mlx and HAS_MLX_WHISPER:
+            logger.info('MLX Whisper backend used.')
             if self._resolved_model_path is not None:
-                mlx_model = str(self._resolved_model_path)
+                mlx_model_path = str(self._resolved_model_path)
             else:
-                mlx_model = mlx_model_mapping.get(self.model_name)
-            if not mlx_model:
+                mlx_model_path = mlx_model_mapping.get(self.model_name)
+            if not mlx_model_path:
                 raise FileNotFoundError(
                     f"MLX Whisper backend requested but no compatible weights found for model '{self.model_name}'."
                 )
-            self.mlx_encoder = load_mlx_encoder(path_or_hf_repo=mlx_model)
+            self.mlx_model = load_mlx_model(path_or_hf_repo=mlx_model_path)
+            self._warmup_mlx_model()
+        elif self.encoder_backend == "mlx-whisper":
+            # hybrid mode: mlx encoder + pytorch decoder
+            logger.info('SimulStreaming will use MLX Whisper encoder with PyTorch decoder.')
+            if self._resolved_model_path is not None:
+                mlx_model_path = str(self._resolved_model_path)
+            else:
+                mlx_model_path = mlx_model_mapping.get(self.model_name)
+            if not mlx_model_path:
+                raise FileNotFoundError(
+                    f"MLX Whisper backend requested but no compatible weights found for model '{self.model_name}'."
+                )
+            self.mlx_encoder = load_mlx_encoder(path_or_hf_repo=mlx_model_path)
+            self.shared_model = self.load_model()
         elif self.encoder_backend == "faster-whisper":
-            print('Simulstreaming will use Faster Whisper for the encoder.')
+            print('SimulStreaming will use Faster Whisper for the encoder.')
             if self._resolved_model_path is not None:
                 fw_model = str(self._resolved_model_path)
             else:
@@ -253,8 +269,20 @@ class SimulStreamingASR():
                 device='auto',
                 compute_type='auto',
             )
-
-        self.models = [self.load_model() for i in range(self.preload_model_count)]
+            self.shared_model = self.load_model()
+        else:
+            self.shared_model = self.load_model()
+    
+    def _warmup_mlx_model(self):
+        """Warmup the full MLX model."""
+        warmup_audio = load_file(self.warmup_file)
+        if warmup_audio is not None:
+            temp_model = MLXAlignAtt(
+                cfg=self.cfg,
+                mlx_model=self.mlx_model,
+            )
+            temp_model.warmup(warmup_audio)
+            logger.info("Full MLX model warmed up successfully")
 
 
     def _resolve_encoder_backend(self, preferred_backend, compatible_whisper_mlx, compatible_faster_whisper):
@@ -298,16 +326,19 @@ class SimulStreamingASR():
         return True
 
     def load_model(self):
+        model_ref = str(self._resolved_model_path) if self._resolved_model_path else self.model_name
+        lora_path = getattr(self, 'lora_path', None)
         whisper_model = load_model(
-            name=self.pytorch_path if self.pytorch_path else self.model_name,
-            download_root=self.model_path,
+            name=model_ref,
+            download_root=getattr(self, 'model_cache_dir', None),
             decoder_only=self.fast_encoder,
-            custom_alignment_heads=self.custom_alignment_heads
-            )
+            custom_alignment_heads=self.custom_alignment_heads,
+            lora_path=lora_path,
+        )
         warmup_audio = load_file(self.warmup_file)
         if warmup_audio is not None:
             warmup_audio = torch.from_numpy(warmup_audio).float()
-            if self.fast_encoder:                
+            if self.fast_encoder:
                 temp_model = AlignAtt(
                     cfg=self.cfg,
                     loaded_model=whisper_model,
@@ -315,32 +346,14 @@ class SimulStreamingASR():
                     fw_encoder=self.fw_encoder,
                 )
                 temp_model.warmup(warmup_audio)
-                temp_model.remove_hooks()
             else:
-                # For standard encoder, use the original transcribe warmup
-                warmup_audio = load_file(self.warmup_file)
                 whisper_model.transcribe(warmup_audio, language=self.lan if self.lan != 'auto' else None)
         return whisper_model
-    
-    def get_new_model_instance(self):
-        """
-        SimulStreaming cannot share the same backend because it uses global forward hooks on the attention layers.
-        Therefore, each user requires a separate model instance, which can be memory-intensive. To maintain speed, we preload the models into memory.
-        """
-        if len(self.models) == 0:
-            self.models.append(self.load_model())
-        new_model = self.models.pop()
-        return new_model
-        # self.models[0]
-
-    def new_model_to_stack(self):
-        self.models.append(self.load_model())
-        
 
     def set_translate_task(self):
         """Set up translation task."""
         if self.cfg.language == 'auto':
-            raise Exception('Translation cannot be done with language = auto')
+            raise ValueError('Translation cannot be done with language = auto')
         return tokenizer.get_tokenizer(
             multilingual=True,
             language=self.cfg.language,

whisperlivekit/simul_whisper/beam.py

@@ -1,17 +1,32 @@
+from torch import Tensor
+
 from whisperlivekit.whisper.decoding import PyTorchInference
 
-# extention of PyTorchInference for beam search
-class BeamPyTorchInference(PyTorchInference):
 
-    def _kv_modules(self):
-        key_modules = [block.attn.key.cache_id for block in self.model.decoder.blocks]
-        value_modules = [block.attn.value.cache_id for block in self.model.decoder.blocks]
-        return key_modules + value_modules
+class BeamPyTorchInference(PyTorchInference):
+    """Extension of PyTorchInference for beam search with cross-attention support."""
+
+    def _kv_cache_ids(self):
+        """Get cache_id strings for self-attention key/value modules."""
+        key_ids = [block.attn.key_cache_id for block in self.model.decoder.blocks]
+        value_ids = [block.attn.value_cache_id for block in self.model.decoder.blocks]
+        return key_ids + value_ids
 
     def rearrange_kv_cache(self, source_indices):
         if source_indices != list(range(len(source_indices))):
-            for module_cache_id in self._kv_modules():
-                self.kv_cache[module_cache_id] = self.kv_cache[module_cache_id][source_indices].detach()
-    from torch import Tensor
-    def logits(self, tokens: Tensor, audio_features: Tensor) -> Tensor:
-        return self.model.decoder(tokens, audio_features, kv_cache=self.kv_cache)
+            for cache_id in self._kv_cache_ids():
+                if cache_id in self.kv_cache:
+                    self.kv_cache[cache_id] = self.kv_cache[cache_id][source_indices].detach()
+
+    def logits(
+        self, 
+        tokens: Tensor, 
+        audio_features: Tensor,
+        return_cross_attn: bool = False,
+    ):
+        """Get logits, optionally returning cross-attention weights."""
+        return self.model.decoder(
+            tokens, audio_features, 
+            kv_cache=self.kv_cache,
+            return_cross_attn=return_cross_attn,
+        )

whisperlivekit/simul_whisper/config.py

@@ -1,6 +1,7 @@
 from dataclasses import dataclass, field
 from typing import Literal
 
+
 @dataclass
 class AlignAttConfig():
     eval_data_path: str = "tmp"

whisperlivekit/simul_whisper/decoder_state.py

@@ -0,0 +1,97 @@
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Tuple
+import torch
+
+
+@dataclass
+class DecoderState:
+
+    kv_cache: Dict[str, torch.Tensor] = field(default_factory=dict)
+    
+    tokenizer: Any = None
+    detected_language: Optional[str] = None
+    reset_tokenizer_to_auto_next_call: bool = False
+    
+    tokens: List[torch.Tensor] = field(default_factory=list)
+    initial_tokens: Optional[torch.Tensor] = None
+    initial_token_length: int = 0
+    sot_index: int = 0
+    
+    align_source: Dict[int, List[Tuple[int, int]]] = field(default_factory=dict)
+    num_align_heads: int = 0
+    
+    segments: List[torch.Tensor] = field(default_factory=list)
+    
+    context: Any = None
+    
+    pending_incomplete_tokens: List[int] = field(default_factory=list)
+    pending_retries: int = 0
+
+    global_time_offset: float = 0.0
+    cumulative_time_offset: float = 0.0
+    first_timestamp: Optional[float] = None
+    last_attend_frame: int = 0
+    
+    speaker: int = -1
+    log_segments: int = 0
+    
+    CIFLinear: Optional[torch.nn.Module] = None
+    always_fire: bool = False
+    never_fire: bool = False
+    
+    suppress_tokens_fn: Any = None
+    
+    token_decoder: Any = None
+    decoder_type: str = "greedy"
+    
+    inference: Any = None
+    
+    def clean_cache(self):
+        """Clean the kv_cache after each inference step."""
+        # Explicitly delete tensor references to free GPU memory
+        if self.kv_cache:
+            for key in list(self.kv_cache.keys()):
+                tensor = self.kv_cache.pop(key, None)
+                if tensor is not None:
+                    del tensor
+
+        # Clear the dict
+        self.kv_cache.clear()
+
+        # Force GPU cache cleanup (only if CUDA is available)
+        import torch
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+
+        if self.decoder_type == "beam" and self.inference is not None:
+            # Create NEW dict instead of sharing reference
+            self.inference.kv_cache = {}
+            if self.token_decoder is not None:
+                self.token_decoder.reset()
+    
+    def reset(self, rewind_threshold: int = 200):
+        """
+        Reset transient state for a new segment.
+        
+        Args:
+            rewind_threshold: Value for resetting last_attend_frame
+        """
+        self.last_attend_frame = -rewind_threshold
+        self.cumulative_time_offset = 0.0
+        self.pending_incomplete_tokens = []
+        self.pending_retries = 0
+        self.log_segments += 1
+
+    def full_reset(self, rewind_threshold: int = 200):
+        """
+        Full reset including audio segments and tokens.
+        
+        Args:
+            rewind_threshold: Value for resetting last_attend_frame
+        """
+        self.reset(rewind_threshold)
+        self.segments = []
+        self.tokens = []
+        self.kv_cache = {}
+        self.first_timestamp = None
+

whisperlivekit/simul_whisper/mlx/__init__.py

@@ -0,0 +1,11 @@
+from .decoder_state import MLXDecoderState
+from .decoders import MLXBeamSearchDecoder, MLXGreedyDecoder, MLXInference
+from .simul_whisper import MLXAlignAtt
+
+__all__ = [
+    "MLXAlignAtt",
+    "MLXBeamSearchDecoder",
+    "MLXDecoderState",
+    "MLXGreedyDecoder",
+    "MLXInference",
+]

whisperlivekit/simul_whisper/mlx/decoder_state.py

@@ -0,0 +1,78 @@
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Tuple
+
+import mlx.core as mx
+import numpy as np
+
+
+@dataclass
+class MLXDecoderState:
+    """
+    mlx kv cache format: List of ((k, v), (cross_k, cross_v)) tuples per layer,
+    where each element is a tuple of mx.arrays.
+    """
+
+    kv_cache: Optional[List[Tuple[Tuple[mx.array, mx.array], Tuple[mx.array, mx.array]]]] = None
+    
+    tokenizer: Any = None
+    detected_language: Optional[str] = None
+    reset_tokenizer_to_auto_next_call: bool = False
+    
+    tokens: List[mx.array] = field(default_factory=list)
+    initial_tokens: Optional[mx.array] = None
+    initial_token_length: int = 0
+    sot_index: int = 0    
+    align_source: Dict[int, List[Tuple[int, int]]] = field(default_factory=dict)
+    num_align_heads: int = 0    
+    segments: List[np.ndarray] = field(default_factory=list)
+    
+    context: Any = None
+    
+    pending_incomplete_tokens: List[int] = field(default_factory=list)
+    pending_retries: int = 0
+
+    global_time_offset: float = 0.0
+    cumulative_time_offset: float = 0.0
+    first_timestamp: Optional[float] = None
+    last_attend_frame: int = 0
+    
+    speaker: int = -1
+    log_segments: int = 0    
+    cif_weights: Optional[mx.array] = None
+    always_fire: bool = False
+    never_fire: bool = False
+    
+    suppress_tokens: Optional[Tuple[int, ...]] = None
+    
+    token_decoder: Any = None
+    decoder_type: str = "greedy"
+    
+    inference: Any = None
+    
+    def clean_cache(self):
+        self.kv_cache = None
+        if self.decoder_type == "beam" and self.inference is not None:
+            self.inference.kv_cache = None
+            if self.token_decoder is not None:
+                self.token_decoder.reset()
+    
+    def reset(self, rewind_threshold: int = 200):
+        self.last_attend_frame = -rewind_threshold
+        self.cumulative_time_offset = 0.0
+        self.pending_incomplete_tokens = []
+        self.pending_retries = 0
+        self.log_segments += 1
+
+    def full_reset(self, rewind_threshold: int = 200):
+        """
+        Full reset including audio segments and tokens.
+        
+        Args:
+            rewind_threshold: Value for resetting last_attend_frame
+        """
+        self.reset(rewind_threshold)
+        self.segments = []
+        self.tokens = []
+        self.kv_cache = None
+        self.first_timestamp = None
+

whisperlivekit/simul_whisper/mlx/decoders.py

@@ -0,0 +1,219 @@
+"""
+MLX-native token decoders for streaming ASR.
+"""
+from typing import Any, Dict, List, Optional, Tuple
+
+import mlx.core as mx
+import numpy as np
+
+
+class MLXGreedyDecoder:
+    """Greedy decoder using MLX operations."""
+    
+    def __init__(self, temperature: float, eot: int):
+        self.temperature = temperature
+        self.eot = eot
+
+    def update(
+        self, tokens: mx.array, logits: mx.array, sum_logprobs: mx.array
+    ) -> Tuple[mx.array, bool]:
+        """
+        Update tokens with next predicted token.
+        
+        Args:
+            tokens: Current token sequence, shape (batch, seq_len)
+            logits: Logits for next token, shape (batch, vocab_size)
+            sum_logprobs: Cumulative log probabilities, shape (batch,)
+            
+        Returns:
+            Updated tokens and completion flag
+        """
+        if self.temperature == 0:
+            next_tokens = mx.argmax(logits, axis=-1)
+        else:
+            probs = mx.softmax(logits / self.temperature, axis=-1)
+            next_tokens = mx.random.categorical(mx.log(probs + 1e-10))
+        
+        logprobs = mx.softmax(logits, axis=-1)
+        logprobs = mx.log(logprobs + 1e-10)        
+        batch_size = logprobs.shape[0]
+        current_logprobs = logprobs[mx.arange(batch_size), next_tokens]        
+        mask = (tokens[:, -1] != self.eot).astype(mx.float32)
+        sum_logprobs = sum_logprobs + current_logprobs * mask        
+        eot_mask = (tokens[:, -1] == self.eot)
+        next_tokens = mx.where(eot_mask, mx.array(self.eot), next_tokens)        
+        tokens = mx.concatenate([tokens, next_tokens[:, None]], axis=1)        
+        completed = bool(mx.all(tokens[:, -1] == self.eot))
+        
+        return tokens, completed
+
+    def finalize(self, tokens: mx.array, sum_logprobs: mx.array):
+        """Finalize decoding by ensuring EOT at end."""
+        eot_column = mx.full((tokens.shape[0], 1), self.eot, dtype=tokens.dtype)
+        tokens = mx.concatenate([tokens, eot_column], axis=1)
+        return tokens, sum_logprobs.tolist()
+
+
+class MLXBeamSearchDecoder:
+    """Beam search decoder using MLX operations."""
+    
+    def __init__(
+        self,
+        beam_size: int,
+        eot: int,
+        inference: Any,
+        patience: Optional[float] = None,
+    ):
+        self.beam_size = beam_size
+        self.eot = eot
+        self.inference = inference
+        self.patience = patience or 1.0
+        self.max_candidates: int = round(beam_size * self.patience)
+        self.finished_sequences: Optional[List[Dict]] = None
+
+        assert (
+            self.max_candidates > 0
+        ), f"Invalid beam size ({beam_size}) or patience ({patience})"
+
+    def reset(self):
+        """Reset finished sequences for new segment."""
+        self.finished_sequences = None
+
+    def update(
+        self, tokens: mx.array, logits: mx.array, sum_logprobs: mx.array
+    ) -> Tuple[mx.array, bool]:
+        """
+        Update tokens using beam search.
+        
+        Args:
+            tokens: Current token sequences, shape (batch * beam_size, seq_len)
+            logits: Logits for next token, shape (batch * beam_size, vocab_size)
+            sum_logprobs: Cumulative log probabilities, shape (batch * beam_size,)
+            
+        Returns:
+            Updated tokens and completion flag
+        """
+        if tokens.shape[0] % self.beam_size != 0:
+            raise ValueError(f"{tokens.shape}[0] % {self.beam_size} != 0")
+
+        n_audio = tokens.shape[0] // self.beam_size
+        if self.finished_sequences is None:
+            self.finished_sequences = [{} for _ in range(n_audio)]
+        logprobs = mx.softmax(logits, axis=-1)
+        logprobs = mx.log(logprobs + 1e-10)        
+        logprobs_np = np.array(logprobs)
+        tokens_np = np.array(tokens)
+        sum_logprobs_np = np.array(sum_logprobs)
+        
+        next_tokens, source_indices, finished_sequences = [], [], []
+        new_sum_logprobs = []
+        
+        for i in range(n_audio):
+            scores, sources, finished = {}, {}, {}
+            for j in range(self.beam_size):
+                idx = i * self.beam_size + j
+                prefix = tokens_np[idx].tolist()                
+                top_k_indices = np.argsort(logprobs_np[idx])[-self.beam_size - 1:][::-1]
+                
+                for token_idx in top_k_indices:
+                    logprob = logprobs_np[idx, token_idx]
+                    new_logprob = sum_logprobs_np[idx] + logprob
+                    sequence = tuple(prefix + [int(token_idx)])
+                    scores[sequence] = new_logprob
+                    sources[sequence] = idx
+            saved = 0
+            for sequence in sorted(scores, key=scores.get, reverse=True):
+                if sequence[-1] == self.eot:
+                    finished[sequence] = scores[sequence]
+                else:
+                    new_sum_logprobs.append(scores[sequence])
+                    next_tokens.append(sequence)
+                    source_indices.append(sources[sequence])
+
+                    saved += 1
+                    if saved == self.beam_size:
+                        break
+
+            finished_sequences.append(finished)
+        tokens = mx.array(np.array(next_tokens, dtype=np.int32))
+        sum_logprobs = mx.array(np.array(new_sum_logprobs, dtype=np.float32))        
+        self.inference.rearrange_kv_cache(source_indices)
+        assert len(self.finished_sequences) == len(finished_sequences)
+        for previously_finished, newly_finished in zip(
+            self.finished_sequences, finished_sequences
+        ):
+            for seq in sorted(newly_finished, key=newly_finished.get, reverse=True):
+                if len(previously_finished) >= self.max_candidates:
+                    break
+                previously_finished[seq] = newly_finished[seq]
+        completed = all(
+            len(sequences) >= self.max_candidates
+            for sequences in self.finished_sequences
+        )
+        
+        return tokens, completed
+
+    def finalize(self, preceding_tokens: mx.array, sum_logprobs: mx.array):
+        """Finalize beam search by selecting best sequences."""
+        preceding_tokens_np = np.array(preceding_tokens)
+        sum_logprobs_np = np.array(sum_logprobs)
+        
+        n_audio = preceding_tokens_np.shape[0] // self.beam_size
+        tokens_list: List[List[int]] = [[] for _ in range(n_audio)]
+        sum_logprobs_list: List[float] = [0.0] * n_audio
+
+        for i, sequences in enumerate(self.finished_sequences):
+            if sequences:
+                best_seq = max(sequences, key=sequences.get)
+                tokens_list[i] = list(best_seq)
+                sum_logprobs_list[i] = sequences[best_seq]
+            else:
+                idx = i * self.beam_size
+                tokens_list[i] = preceding_tokens_np[idx].tolist() + [self.eot]
+                sum_logprobs_list[i] = float(sum_logprobs_np[idx])
+        max_len = max(len(t) for t in tokens_list)
+        for i, t in enumerate(tokens_list):
+            tokens_list[i] = t + [self.eot] * (max_len - len(t))
+
+        tokens = mx.array(np.array(tokens_list, dtype=np.int32))
+        return tokens, sum_logprobs_list
+
+
+class MLXInference:
+    """MLX inference wrapper for beam search KV cache management."""
+    
+    def __init__(self, model, initial_token_length: int):
+        self.model = model
+        self.initial_token_length = initial_token_length
+        self.kv_cache = None
+    
+    def rearrange_kv_cache(self, source_indices: List[int]):
+        """Rearrange KV cache based on beam search source indices."""
+        if self.kv_cache is None:
+            return
+            
+        if source_indices == list(range(len(source_indices))):
+            return
+        
+        source_indices_mx = mx.array(source_indices, dtype=mx.int32)
+        
+        new_cache = []
+        for layer_cache in self.kv_cache:
+            (k, v), (cross_k, cross_v) = layer_cache            
+            new_k = k[source_indices_mx]
+            new_v = v[source_indices_mx]
+            new_cache.append(((new_k, new_v), (cross_k, cross_v)))
+        
+        self.kv_cache = new_cache
+    
+    def logits(
+        self, 
+        tokens: mx.array, 
+        audio_features: mx.array,
+    ) -> Tuple[mx.array, List]:
+        """Get logits from decoder with KV cache."""
+        logits, self.kv_cache, cross_qk = self.model.decoder(
+            tokens, audio_features, kv_cache=self.kv_cache
+        )
+        return logits, cross_qk
+

whisperlivekit/simul_whisper/mlx/simul_whisper.py

@@ -0,0 +1,421 @@
+"""MLX whisper AlignAtt streaming decoder."""
+import logging
+from typing import Any, List, Tuple
+
+import mlx.core as mx
+import numpy as np
+
+from mlx_whisper.audio import log_mel_spectrogram as mlx_log_mel_spectrogram
+from mlx_whisper.transcribe import pad_or_trim as mlx_pad_or_trim
+
+from whisperlivekit.whisper.audio import N_FRAMES, N_SAMPLES, TOKENS_PER_SECOND
+
+from ..align_att_base import DEC_PAD, AlignAttBase
+from ..config import AlignAttConfig
+from .decoder_state import MLXDecoderState
+from .decoders import MLXBeamSearchDecoder, MLXGreedyDecoder, MLXInference
+
+
+logger = logging.getLogger(__name__)
+
+
+class MLXTokenBuffer:
+    """Token buffer for MLX-based decoding."""
+
+    def __init__(self, text="", tokenizer=None, prefix_token_ids=None):
+        self.text = text
+        self.prefix_token_ids = prefix_token_ids or []
+        self.tokenizer = tokenizer
+        self.pending_token_ids = []
+
+    def as_token_ids(self, tokenizer=None):
+        if tokenizer is None:
+            tokenizer = self.tokenizer
+        if tokenizer is None:
+            raise ValueError("Tokenizer is not set.")
+        return self.prefix_token_ids + tokenizer.encode(self.text)
+
+    def as_mlx_array(self) -> mx.array:
+        tok_ids = self.as_token_ids()
+        return mx.array([tok_ids], dtype=mx.int32)
+
+    def as_mlx_array_beam(self, beam: int) -> mx.array:
+        t = self.as_mlx_array()
+        return mx.repeat(t, beam, axis=0)
+
+    def as_text(self):
+        return self.text
+
+    @staticmethod
+    def empty(*a, **kw):
+        return MLXTokenBuffer(*a, **kw)
+
+    @staticmethod
+    def from_text(text, *a, **kw):
+        return MLXTokenBuffer(*a, text=text, **kw)
+
+    def is_empty(self):
+        return self.text is None or self.text == ""
+
+    def trim_words(self, num=1, after=0):
+        tokenizer = self.tokenizer
+        assert tokenizer is not None, "Tokenizer is not set."
+        ids = tokenizer.encode(self.text[after:])
+        words, wids = self.tokenizer.split_to_word_tokens(ids)
+        if not words:
+            return 0
+        self.text = self.text[:after] + "".join(words[num:])
+        return sum(len(wi) for wi in wids[:num])
+
+    def append_token_ids(self, token_ids):
+        tokenizer = self.tokenizer
+        assert tokenizer is not None, "Tokenizer is not set."
+        all_tokens = self.pending_token_ids + token_ids
+        decoded = tokenizer.decode(all_tokens)
+        replacement_char = "\ufffd"
+        if replacement_char in decoded:
+            if len(all_tokens) > 1:
+                decoded_partial = tokenizer.decode(all_tokens[:-1])
+                if replacement_char not in decoded_partial:
+                    self.text += decoded_partial
+                    self.pending_token_ids = [all_tokens[-1]]
+                else:
+                    self.pending_token_ids = all_tokens
+            else:
+                self.pending_token_ids = all_tokens
+        else:
+            self.text += decoded
+            self.pending_token_ids = []
+
+
+def mlx_median_filter(x: mx.array, filter_width: int) -> mx.array:
+    """Apply median filter along the last axis."""
+    if filter_width <= 1:
+        return x
+    pad_width = filter_width // 2
+    shape = x.shape
+    left_pad = mx.repeat(x[..., :1], pad_width, axis=-1)
+    right_pad = mx.repeat(x[..., -1:], pad_width, axis=-1)
+    x_padded = mx.concatenate([left_pad, x, right_pad], axis=-1)
+    result = []
+    for i in range(shape[-1]):
+        window = x_padded[..., i:i + filter_width]
+        sorted_window = mx.sort(window, axis=-1)
+        median_val = sorted_window[..., filter_width // 2:filter_width // 2 + 1]
+        result.append(median_val)
+    return mx.concatenate(result, axis=-1)
+
+
+class MLXAlignAtt(AlignAttBase):
+    """
+    MLX-native Alignment-based Attention decoder for SimulStreaming.
+
+    Runs entirely on MLX, with no PyTorch dependencies for inference.
+    """
+
+    def __init__(
+        self,
+        cfg: AlignAttConfig,
+        mlx_model: Any,
+    ) -> None:
+        # Common init (sets self.model, self.cfg, decode_options, etc.)
+        self._base_init(cfg, mlx_model)
+        logger.info(f"MLX Model dimensions: {self.model.dims}")
+
+        # Per-session state
+        self.state = MLXDecoderState()
+        self._init_state(cfg)
+
+    def _init_state(self, cfg: AlignAttConfig):
+        self._init_state_common(cfg)
+
+        # CIF: MLX doesn't support CIF checkpoint loading
+        if cfg.cif_ckpt_path is None or not cfg.cif_ckpt_path:
+            if cfg.never_fire:
+                self.state.never_fire = True
+                self.state.always_fire = False
+            else:
+                self.state.always_fire = True
+                self.state.never_fire = False
+        else:
+            logger.warning(
+                "CIF checkpoint provided but MLX CIF not implemented. "
+                "Using always_fire=True"
+            )
+            self.state.always_fire = True
+            self.state.never_fire = cfg.never_fire
+
+        self._build_alignment_source()
+
+        # Suppress tokens
+        suppress_tokens = [
+            self.tokenizer.transcribe, self.tokenizer.translate,
+            self.tokenizer.sot, self.tokenizer.sot_prev,
+            self.tokenizer.sot_lm, self.tokenizer.no_timestamps,
+        ] + list(self.tokenizer.all_language_tokens)
+        if self.tokenizer.no_speech is not None:
+            suppress_tokens.append(self.tokenizer.no_speech)
+        self.state.suppress_tokens = tuple(sorted(set(suppress_tokens)))
+        logger.debug(f"Suppress tokens: {self.state.suppress_tokens}")
+
+        self.init_tokens()
+        self.init_context()
+
+        # Decoder type
+        self.state.decoder_type = cfg.decoder_type
+        if cfg.decoder_type == "greedy":
+            logger.info("Using MLX greedy decoder")
+            self.state.token_decoder = MLXGreedyDecoder(0.0, self.tokenizer.eot)
+        elif cfg.decoder_type == "beam":
+            logger.info("Using MLX beam decoder")
+            self.state.inference = MLXInference(
+                self.model, self.state.initial_token_length,
+            )
+            self.state.token_decoder = MLXBeamSearchDecoder(
+                inference=self.state.inference,
+                eot=self.tokenizer.eot,
+                beam_size=cfg.beam_size,
+            )
+
+    def _build_alignment_source(self):
+        """Build alignment source mapping from model's alignment_heads."""
+        self.state.align_source = {}
+        self.state.num_align_heads = 0
+        alignment_heads = self.model.alignment_heads
+        if alignment_heads is None:
+            logger.warning("No alignment heads found in model")
+            return
+        if hasattr(alignment_heads, 'tolist'):
+            heads_list = alignment_heads.tolist()
+        else:
+            heads_list = np.array(alignment_heads).tolist()
+        for layer_rank, head_id in heads_list:
+            layer_rank = int(layer_rank)
+            head_id = int(head_id)
+            heads = self.state.align_source.get(layer_rank, [])
+            heads.append((self.state.num_align_heads, head_id))
+            self.state.align_source[layer_rank] = heads
+            self.state.num_align_heads += 1
+
+    # === Abstract method implementations ===
+
+    def init_tokens(self):
+        logger.debug(f"init tokens, {len(self.state.segments)}")
+        self.state.initial_tokens = mx.array(
+            [self.tokenizer.sot_sequence_including_notimestamps],
+            dtype=mx.int32,
+        )
+        self.state.initial_token_length = self.state.initial_tokens.shape[1]
+        self.state.sot_index = self.tokenizer.sot_sequence.index(self.tokenizer.sot)
+        logger.debug(f"init tokens after, {len(self.state.segments)}")
+        self.state.tokens = [self.state.initial_tokens]
+
+    def init_context(self):
+        kw = {
+            'tokenizer': self.tokenizer,
+            'prefix_token_ids': [self.tokenizer.sot_prev],
+        }
+        self.state.context = MLXTokenBuffer.empty(**kw)
+        if self.cfg.static_init_prompt is not None:
+            self.state.context = MLXTokenBuffer.from_text(self.cfg.static_init_prompt, **kw)
+        if self.cfg.init_prompt is not None:
+            self.state.context.text += self.cfg.init_prompt
+
+    def insert_audio(self, segment=None):
+        if segment is not None:
+            if hasattr(segment, 'numpy'):
+                segment = segment.numpy()
+            self.state.segments.append(segment)
+        removed_len = 0
+        segments_len = self.segments_len()
+        while len(self.state.segments) > 1 and segments_len > self.cfg.audio_max_len:
+            removed_len = self.state.segments[0].shape[0] / 16000
+            segments_len -= removed_len
+            self.state.last_attend_frame -= int(TOKENS_PER_SECOND * removed_len)
+            self.state.cumulative_time_offset += removed_len
+            self.state.segments = self.state.segments[1:]
+            logger.debug(
+                f"remove segments: {len(self.state.segments)} {len(self.state.tokens)}, "
+                f"cumulative offset: {self.state.cumulative_time_offset:.2f}s"
+            )
+            if len(self.state.tokens) > 1:
+                token_list = np.array(self.state.tokens[1][0, :]).tolist()
+                self.state.context.append_token_ids(token_list)
+                self.state.tokens = [self.state.initial_tokens] + self.state.tokens[2:]
+        return removed_len
+
+    def _current_tokens(self) -> mx.array:
+        toks = self.state.tokens
+        if toks[0].shape[0] == 1:
+            toks[0] = mx.repeat(toks[0], self.cfg.beam_size, axis=0)
+        if not self.state.context.is_empty():
+            context_toks = self.state.context.as_mlx_array_beam(self.cfg.beam_size)
+            toks = [context_toks] + toks
+        if len(toks) > 1:
+            current_tokens = mx.concatenate(toks, axis=1)
+        else:
+            current_tokens = toks[0]
+        logger.debug("debug print current_tokens:")
+        self.debug_print_tokens(current_tokens)
+        return current_tokens
+
+    def fire_at_boundary(self, chunked_encoder_feature: mx.array) -> bool:
+        if self.state.always_fire:
+            return True
+        if self.state.never_fire:
+            return False
+        return True  # MLX CIF not implemented
+
+    def lang_id(self, encoder_features: mx.array) -> Tuple[mx.array, List[dict]]:
+        n_audio = encoder_features.shape[0]
+        x = mx.array([[self.tokenizer.sot]] * n_audio, dtype=mx.int32)
+        logits, _, _ = self.model.decoder(x, encoder_features, kv_cache=None)
+        logits = logits[:, 0]
+
+        mask = mx.ones(logits.shape[-1], dtype=mx.bool_)
+        language_token_indices = mx.array(
+            list(self.tokenizer.all_language_tokens), dtype=mx.int32,
+        )
+        mask = mask.at[language_token_indices].add(False)
+        logits = mx.where(mask, mx.array(-float('inf')), logits)
+
+        language_tokens = mx.argmax(logits, axis=-1)
+        language_token_probs = mx.softmax(logits, axis=-1)
+        probs_np = np.array(language_token_probs)
+        language_probs = [
+            {
+                c: float(probs_np[i, j])
+                for j, c in zip(
+                    self.tokenizer.all_language_tokens,
+                    self.tokenizer.all_language_codes,
+                )
+            }
+            for i in range(n_audio)
+        ]
+        self._clean_cache()
+        return language_tokens, language_probs
+
+    def _concat_segments(self):
+        if len(self.state.segments) > 1:
+            return np.concatenate(self.state.segments, axis=0)
+        return self.state.segments[0]
+
+    def _encode(self, input_segments):
+        mlx_mel_padded = mlx_log_mel_spectrogram(
+            audio=input_segments,
+            n_mels=self.model.dims.n_mels,
+            padding=N_SAMPLES,
+        )
+        mlx_mel = mlx_pad_or_trim(mlx_mel_padded, N_FRAMES, axis=-2)
+        encoder_feature = self.model.encoder(mlx_mel[None])
+        content_mel_len = int((mlx_mel_padded.shape[0] - mlx_mel.shape[0]) / 2)
+        return encoder_feature, content_mel_len
+
+    def _init_sum_logprobs(self):
+        return mx.zeros((self.cfg.beam_size,), dtype=mx.float32)
+
+    def _get_logits_and_cross_attn(self, tokens, encoder_feature):
+        if self.state.decoder_type == "greedy":
+            logits, self.state.kv_cache, cross_qk = self.model.decoder(
+                tokens, encoder_feature, kv_cache=self.state.kv_cache,
+            )
+            return logits, cross_qk
+        else:
+            return self.state.inference.logits(tokens, encoder_feature)
+
+    def _check_no_speech(self, logits):
+        if self.tokenizer.no_speech is not None:
+            probs_at_sot = mx.softmax(logits[:, self.state.sot_index, :], axis=-1)
+            no_speech_probs = np.array(
+                probs_at_sot[:, self.tokenizer.no_speech],
+            ).tolist()
+            if no_speech_probs[0] > self.cfg.nonspeech_prob:
+                logger.info("no speech, stop")
+                return True
+        return False
+
+    def _suppress_blank_tokens(self, logits):
+        blank_tokens = self.tokenizer.encode(" ") + [self.tokenizer.eot]
+        logits = logits.at[:, blank_tokens].add(-float('inf'))
+        return logits
+
+    def _apply_token_suppression(self, logits):
+        if self.state.suppress_tokens:
+            suppress_indices = mx.array(
+                list(self.state.suppress_tokens), dtype=mx.int32,
+            )
+            logits = logits.at[:, suppress_indices].add(-float('inf'))
+        return logits
+
+    def _update_tokens(self, current_tokens, logits, sum_logprobs):
+        return self.state.token_decoder.update(current_tokens, logits, sum_logprobs)
+
+    def _process_cross_attention(
+        self, cross_attns: List, content_mel_len: int,
+    ) -> mx.array:
+        attn_of_alignment_heads = [[] for _ in range(self.state.num_align_heads)]
+        num_decoder_layers = self.num_decoder_layers
+
+        if cross_attns and isinstance(cross_attns[0], list):
+            flattened_attns = [attn for layer_list in cross_attns for attn in layer_list]
+        else:
+            flattened_attns = cross_attns
+
+        for idx, attn_mat in enumerate(flattened_attns):
+            if attn_mat is None:
+                continue
+            layer_rank = idx % num_decoder_layers
+            align_heads_in_layer = self.state.align_source.get(layer_rank, [])
+            if not align_heads_in_layer:
+                continue
+            attn_mat = mx.softmax(attn_mat, axis=-1)
+            for align_head_rank, head_id in align_heads_in_layer:
+                if self.cfg.beam_size == 1:
+                    if attn_mat.ndim == 4:
+                        a = attn_mat[0, head_id, :, :]
+                    else:
+                        a = attn_mat[head_id, :, :]
+                    a = a[None, :, :]
+                else:
+                    a = attn_mat[:, head_id, :, :]
+                attn_of_alignment_heads[align_head_rank].append(a)
+
+        tmp = []
+        for mat in attn_of_alignment_heads:
+            if mat:
+                tmp.append(mx.concatenate(mat, axis=1))
+        if not tmp:
+            return mx.zeros((self.cfg.beam_size, 1, content_mel_len))
+
+        attn_of_alignment_heads = mx.stack(tmp, axis=1)
+        std = mx.std(attn_of_alignment_heads, axis=-2, keepdims=True)
+        mean = mx.mean(attn_of_alignment_heads, axis=-2, keepdims=True)
+        attn_of_alignment_heads = (attn_of_alignment_heads - mean) / (std + 1e-8)
+        attn_of_alignment_heads = mlx_median_filter(attn_of_alignment_heads, 7)
+        attn_of_alignment_heads = mx.mean(attn_of_alignment_heads, axis=1)
+        attn_of_alignment_heads = attn_of_alignment_heads[:, :, :content_mel_len]
+        mx.eval(attn_of_alignment_heads)
+        return attn_of_alignment_heads
+
+    def _get_attended_frames(self, attn):
+        most_attended_frames = mx.argmax(attn[:, -1, :], axis=-1)
+        frames_np = np.array(most_attended_frames)
+        return frames_np.tolist(), int(frames_np[0])
+
+    def _is_special_token(self, current_tokens):
+        return int(np.array(current_tokens[0, -2])) >= DEC_PAD
+
+    def _rewind_tokens(self):
+        if len(self.state.tokens) > 0:
+            return mx.concatenate(self.state.tokens, axis=1)
+        return self.state.tokens[0]
+
+    def _tokens_to_list(self, current_tokens, start_col):
+        return np.array(current_tokens[0, start_col:]).tolist()
+
+    def _make_new_tokens_tensor(self, hypothesis):
+        new_tokens = mx.array([hypothesis], dtype=mx.int32)
+        return mx.repeat(new_tokens, self.cfg.beam_size, axis=0)
+
+    def _evaluate(self, tensor):
+        mx.eval(tensor)

whisperlivekit/simul_whisper/mlx_encoder.py

@@ -5,24 +5,11 @@ import mlx.core as mx
 import mlx.nn as nn
 from huggingface_hub import snapshot_download
 from mlx.utils import tree_unflatten
-
 from mlx_whisper import whisper
 
-mlx_model_mapping = {
-    "tiny.en": "mlx-community/whisper-tiny.en-mlx",
-    "tiny": "mlx-community/whisper-tiny-mlx",
-    "base.en": "mlx-community/whisper-base.en-mlx",
-    "base": "mlx-community/whisper-base-mlx",
-    "small.en": "mlx-community/whisper-small.en-mlx",
-    "small": "mlx-community/whisper-small-mlx",
-    "medium.en": "mlx-community/whisper-medium.en-mlx",
-    "medium": "mlx-community/whisper-medium-mlx",
-    "large-v1": "mlx-community/whisper-large-v1-mlx",
-    "large-v2": "mlx-community/whisper-large-v2-mlx",
-    "large-v3": "mlx-community/whisper-large-v3-mlx",
-    "large-v3-turbo": "mlx-community/whisper-large-v3-turbo",
-    "large": "mlx-community/whisper-large-mlx",
-}
+from whisperlivekit.model_mapping import MLX_MODEL_MAPPING
+
+mlx_model_mapping = MLX_MODEL_MAPPING
 
 def load_mlx_encoder(
     path_or_hf_repo: str,
@@ -69,4 +56,40 @@ def load_mlx_encoder(
 
     model.update(encoder_weights)
     mx.eval(model.parameters())
+    return model
+
+
+def load_mlx_model(
+    path_or_hf_repo: str,
+    dtype: mx.Dtype = mx.float32,
+) -> whisper.Whisper:
+    model_path = Path(path_or_hf_repo)
+    if not model_path.exists():
+        model_path = Path(snapshot_download(repo_id=path_or_hf_repo))
+
+    with open(str(model_path / "config.json"), "r") as f:
+        config = json.loads(f.read())
+        config.pop("model_type", None)
+        quantization = config.pop("quantization", None)
+
+    model_args = whisper.ModelDimensions(**config)
+
+    wf = model_path / "weights.safetensors"
+    if not wf.exists():
+        wf = model_path / "weights.npz"
+    weights = mx.load(str(wf))
+
+    model = whisper.Whisper(model_args, dtype)
+
+    if quantization is not None:
+        class_predicate = (
+            lambda p, m: isinstance(m, (nn.Linear, nn.Embedding))
+            and f"{p}.scales" in weights
+        )
+        nn.quantize(model, **quantization, class_predicate=class_predicate)
+
+    weights = tree_unflatten(list(weights.items()))
+    
+    model.update(weights)
+    mx.eval(model.parameters())
     return model

whisperlivekit/simul_whisper/simul_whisper.py

@@ -1,33 +1,32 @@
-import os
 import logging
+import os
+from typing import List
 
+import numpy as np
 import torch
 import torch.nn.functional as F
-import numpy as np
 
-from whisperlivekit.whisper import DecodingOptions, tokenizer
-from .config import AlignAttConfig
-from whisperlivekit.timed_objects import ASRToken
-from whisperlivekit.whisper.audio import log_mel_spectrogram, TOKENS_PER_SECOND, pad_or_trim, N_SAMPLES, N_FRAMES
+from whisperlivekit.backend_support import (faster_backend_available,
+                                            mlx_backend_available)
+from whisperlivekit.whisper.audio import (N_FRAMES, N_SAMPLES,
+                                          TOKENS_PER_SECOND,
+                                          log_mel_spectrogram, pad_or_trim)
+from whisperlivekit.whisper.decoding import (BeamSearchDecoder, GreedyDecoder,
+                                             SuppressTokens)
 from whisperlivekit.whisper.timing import median_filter
-from whisperlivekit.whisper.decoding import GreedyDecoder, BeamSearchDecoder, SuppressTokens
+
+from .align_att_base import DEC_PAD, AlignAttBase
 from .beam import BeamPyTorchInference
+from .config import AlignAttConfig
+from .decoder_state import DecoderState
 from .eow_detection import fire_at_boundary, load_cif
-import os
-from time import time
 from .token_buffer import TokenBuffer
-from whisperlivekit.backend_support import (
-    mlx_backend_available,
-    faster_backend_available,
-)
 
-from ..timed_objects import PUNCTUATION_MARKS
-
-DEC_PAD = 50257
 logger = logging.getLogger(__name__)
 
 if mlx_backend_available():
-    from mlx_whisper.audio import log_mel_spectrogram as mlx_log_mel_spectrogram
+    from mlx_whisper.audio import \
+        log_mel_spectrogram as mlx_log_mel_spectrogram
     from mlx_whisper.transcribe import pad_or_trim as mlx_pad_or_trim
 
 if faster_backend_available():
@@ -43,7 +42,10 @@ def load_coreml_encoder():
     except ImportError:
         logger.warning("coremltools is not installed")
         return None
-    COREML_ENCODER_PATH = os.environ.get("MLCORE_ENCODER_PATH", "whisperlivekit/whisper/whisper_encoder.mlpackage")
+    COREML_ENCODER_PATH = os.environ.get(
+        "MLCORE_ENCODER_PATH",
+        "whisperlivekit/whisper/whisper_encoder.mlpackage",
+    )
     _coreml_encoder = MLModel(COREML_ENCODER_PATH)
     spec = _coreml_encoder.get_spec()
     _coreml_input_name = spec.description.input[0].name if spec.description.input else "mel"
@@ -51,252 +53,146 @@ def load_coreml_encoder():
     return _coreml_encoder, _coreml_input_name, _coreml_output_name
 
 
-class AlignAtt:
+class AlignAtt(AlignAttBase):
+    """
+    PyTorch Alignment-based Attention decoder for SimulStreaming.
+
+    Hookless — the model can be shared across multiple sessions,
+    with each session maintaining its own DecoderState.
+    """
+
     def __init__(
-            self, 
-            cfg: AlignAttConfig,
-            loaded_model=None,
-            mlx_encoder=None,
-            fw_encoder=None,
-        ) -> None:
-        self.log_segments = 0
-        
-        self.model = loaded_model
+        self,
+        cfg: AlignAttConfig,
+        loaded_model=None,
+        mlx_encoder=None,
+        fw_encoder=None,
+    ) -> None:
         self.mlx_encoder = mlx_encoder
-        self.fw_encoder = fw_encoder            
+        self.fw_encoder = fw_encoder
         if fw_encoder:
-            self.fw_feature_extractor = FeatureExtractor(feature_size=self.model.dims.n_mels)
+            self.fw_feature_extractor = FeatureExtractor(
+                feature_size=loaded_model.dims.n_mels,
+            )
         self.coreml_encoder_tuple = None
         if USE_MLCORE:
             self.coreml_encoder_tuple = load_coreml_encoder()
         self.use_mlcore = self.coreml_encoder_tuple is not None
-        
         self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
-        
+
+        # Common init (sets self.model, self.cfg, decode_options, etc.)
+        self._base_init(cfg, loaded_model)
         logger.info(f"Model dimensions: {self.model.dims}")
-        self.speaker = -1
-        self.decode_options = DecodingOptions(
-            language = cfg.language, 
-            without_timestamps = True,
-            task=cfg.task
+
+        # Per-session state
+        self.state = DecoderState()
+        self._init_state(cfg)
+
+    def _init_state(self, cfg: AlignAttConfig):
+        self._init_state_common(cfg)
+
+        # CIF helpers for end-of-word boundary detection
+        self.state.CIFLinear, self.state.always_fire, self.state.never_fire = load_cif(
+            cfg, n_audio_state=self.model.dims.n_audio_state, device=self.model.device,
         )
-        self.tokenizer_is_multilingual = cfg.tokenizer_is_multilingual
-        self.create_tokenizer(cfg.language if cfg.language != "auto" else None)
-        # self.create_tokenizer('en')
-        self.detected_language = cfg.language if cfg.language != "auto" else None
-        self.global_time_offset = 0.0
-        self.reset_tokenizer_to_auto_next_call = False
-        
-        self.max_text_len = self.model.dims.n_text_ctx
-        self.num_decoder_layers = len(self.model.decoder.blocks)
-        self.cfg = cfg
-        self.l_hooks = []
 
-        # model to detect end-of-word boundary at the end of the segment
-        self.CIFLinear, self.always_fire, self.never_fire = load_cif(cfg,
-                                                                     n_audio_state=self.model.dims.n_audio_state,
-                                                                     device=self.model.device)
-
-        # install hooks to access encoder-decoder attention
-        self.dec_attns = []
-        def layer_hook(module, net_input, net_output):
-            # net_output[1]: B*num_head*token_len*audio_len
-            t = F.softmax(net_output[1], dim=-1)
-            self.dec_attns.append(t.squeeze(0))
-        for b in self.model.decoder.blocks:
-            hook = b.cross_attn.register_forward_hook(layer_hook)
-            self.l_hooks.append(hook)
-        
-        self.kv_cache = {}
-        def kv_hook(module: torch.nn.Linear, _, net_output: torch.Tensor):
-            if module.cache_id not in self.kv_cache or net_output.shape[1] > self.max_text_len:
-                # save as-is, for the first token or cross attention
-                self.kv_cache[module.cache_id] = net_output
-            else:
-                x = self.kv_cache[module.cache_id]
-                self.kv_cache[module.cache_id] = torch.cat([x, net_output], dim=1).detach()
-            return self.kv_cache[module.cache_id] 
-
-        for i,b in enumerate(self.model.decoder.blocks):
-            hooks = [
-                b.attn.key.register_forward_hook(kv_hook),
-                b.attn.value.register_forward_hook(kv_hook),
-                b.cross_attn.key.register_forward_hook(kv_hook),
-                b.cross_attn.value.register_forward_hook(kv_hook),
-            ]
-            self.l_hooks.extend(hooks)
-
-        self.align_source = {}
-        self.num_align_heads = 0
+        # Build alignment source mapping
+        self.state.align_source = {}
+        self.state.num_align_heads = 0
         for layer_rank, head_id in self.model.alignment_heads.indices().T:
             layer_rank = layer_rank.item()
-            heads = self.align_source.get(layer_rank, [])
-            heads.append((self.num_align_heads, head_id.item()))
-            self.align_source[layer_rank] = heads
-            self.num_align_heads += 1
+            heads = self.state.align_source.get(layer_rank, [])
+            heads.append((self.state.num_align_heads, head_id.item()))
+            self.state.align_source[layer_rank] = heads
+            self.state.num_align_heads += 1
 
-
-        # tokens to be suppressed from decoding, to prevent hallucinations
+        # Build suppress tokens function
         suppress_tokens = [
-                self.tokenizer.transcribe,
-                self.tokenizer.translate,
-                self.tokenizer.sot,
-                self.tokenizer.sot_prev,
-                self.tokenizer.sot_lm,
-                # self.tokenizer.eot 
-                self.tokenizer.no_timestamps,  # added by DM
-            ] + list(self.tokenizer.all_language_tokens)  # added by DM
+            self.tokenizer.transcribe, self.tokenizer.translate,
+            self.tokenizer.sot, self.tokenizer.sot_prev,
+            self.tokenizer.sot_lm, self.tokenizer.no_timestamps,
+        ] + list(self.tokenizer.all_language_tokens)
         if self.tokenizer.no_speech is not None:
             suppress_tokens.append(self.tokenizer.no_speech)
-        suppress_tokens =  tuple(sorted(set(suppress_tokens)))
+        suppress_tokens = tuple(sorted(set(suppress_tokens)))
         logger.debug(f"Suppress tokens: {suppress_tokens}")
         sup_tokens = SuppressTokens(suppress_tokens)
-        self.suppress_tokens = lambda logits: sup_tokens.apply(logits, None)
-        # blank tokens are suppresed for new segments near the line 334
+        self.state.suppress_tokens_fn = lambda logits: sup_tokens.apply(logits, None)
 
-        # it's going to be regenerated after lang id
-        self.segments = []
         self.init_tokens()
-        
-        self.last_attend_frame = -self.cfg.rewind_threshold
-        self.cumulative_time_offset = 0.0
-        self.first_timestamp = None
-
-        if self.cfg.max_context_tokens is None:
-            self.max_context_tokens = self.max_text_len
-        else:
-            self.max_context_tokens = self.cfg.max_context_tokens
         self.init_context()
 
-        # decoder type: greedy or beam
+        # Decoder type
+        self.state.decoder_type = cfg.decoder_type
         if cfg.decoder_type == "greedy":
             logger.info("Using greedy decoder")
-            self.token_decoder = GreedyDecoder(0.0, self.tokenizer.eot)
-            self.decoder_type = "greedy"
-
+            self.state.token_decoder = GreedyDecoder(0.0, self.tokenizer.eot)
         elif cfg.decoder_type == "beam":
-            self.decoder_type = "beam"
-            self.inference = BeamPyTorchInference(self.model, self.initial_token_length)
-            self.inference.kv_cache = self.kv_cache
+            logger.info("Using beam decoder")
+            self.state.inference = BeamPyTorchInference(
+                self.model, self.state.initial_token_length,
+            )
+            self.state.inference.kv_cache = self.state.kv_cache
+            self.state.token_decoder = BeamSearchDecoder(
+                inference=self.state.inference,
+                eot=self.tokenizer.eot,
+                beam_size=cfg.beam_size,
+            )
 
-            self.token_decoder = BeamSearchDecoder(inference=self.inference, eot=self.tokenizer.eot, beam_size=cfg.beam_size)
-
-        # Tokens to carry over to next chunk for incomplete UTF-8 characters
-        self.pending_incomplete_tokens = []
-
-    def remove_hooks(self):
-        for hook in self.l_hooks:
-            hook.remove()
-
-    def warmup(self, audio):
-        try:
-            self.insert_audio(audio)
-            self.infer(is_last=True)
-            self.refresh_segment(complete=True)
-            logger.info("Model warmed up successfully")
-        except Exception as e:
-            logger.exception(f"Model warmup failed: {e}")
-
-    def create_tokenizer(self, language=None):
-        self.tokenizer = tokenizer.get_tokenizer(
-            multilingual=self.tokenizer_is_multilingual,  
-            language=language,
-            num_languages=self.model.num_languages,
-            task=self.decode_options.task
-        )
-
-    def init_context(self):
-        kw = {'tokenizer': self.tokenizer, 
-              'device': self.model.device, 
-              'prefix_token_ids': [self.tokenizer.sot_prev]}
-        self.context = TokenBuffer.empty(**kw)
-        if self.cfg.static_init_prompt is not None:
-            self.context = TokenBuffer.from_text(self.cfg.static_init_prompt, **kw)
-        if self.cfg.init_prompt is not None:
-            self.context.text += self.cfg.init_prompt
+    # === Abstract method implementations ===
 
     def init_tokens(self):
-        logger.debug(f"init tokens, {len(self.segments)}")
-        # init tokens (mandatory prompt)
-        self.initial_tokens = torch.tensor(
-            self.tokenizer.sot_sequence_including_notimestamps, 
-            dtype=torch.long, 
-            device=self.model.device).unsqueeze(0)
-        self.initial_token_length = self.initial_tokens.shape[1]
-        self.sot_index = self.tokenizer.sot_sequence.index(self.tokenizer.sot)
-#        self.segments = []
-        logger.debug(f"init tokens after, {len(self.segments)}")
-        self.tokens = [self.initial_tokens]
+        logger.debug(f"init tokens, {len(self.state.segments)}")
+        self.state.initial_tokens = torch.tensor(
+            self.tokenizer.sot_sequence_including_notimestamps,
+            dtype=torch.long, device=self.model.device,
+        ).unsqueeze(0)
+        self.state.initial_token_length = self.state.initial_tokens.shape[1]
+        self.state.sot_index = self.tokenizer.sot_sequence.index(self.tokenizer.sot)
+        logger.debug(f"init tokens after, {len(self.state.segments)}")
+        self.state.tokens = [self.state.initial_tokens]
 
-    def trim_context(self):
-        logger.info("Trimming context")
-        c = len(self.context.as_token_ids()) - len(self.context.prefix_token_ids)
-#        logger.debug(f"c= {len(self.context.as_token_ids())}, {len(self.context.prefix_token_ids)}")
-        logger.info(f"Context text: {self.context.as_text()}")
-#        logger.debug(f"Context tensor: {self.context.as_tensor()}")
-        l = sum(t.shape[1] for t in self.tokens) + c
-#        logger.debug(f"len {l}, c {c}, max_context_tokens {self.max_context_tokens}")
-        if self.cfg.static_init_prompt is None:
-            after = 0
-        else:
-            after = len(self.cfg.static_init_prompt)
-#        logger.debug(f"len {l}, c {c}, max_context_tokens {self.max_context_tokens}")
-        while c > self.max_context_tokens or l > self.max_text_len - 20:
-            t = self.context.trim_words(after=after)
-            l -= t
-            c -= t
-            logger.debug(f"len {l}, c {c}, max_context_tokens {self.max_context_tokens}")
-            if t == 0:
-                break
-#        logger.debug(f"len {l}, c {c}, max_context_tokens {self.max_context_tokens}")
-        logger.info(f"Context after trim: {self.context.text} (len: {l})")
-
-
-    def logits(self, tokens: torch.Tensor, audio_features: torch.Tensor) -> torch.Tensor:
-        if self.cfg.decoder_type == "greedy":
-            logit = self.model.decoder(tokens, audio_features, kv_cache=self.kv_cache)
-        else:
-            logger.debug(f"Logits shape: {tokens.shape}")
-            logit = self.inference.logits(tokens, audio_features)
-        return logit
-    
-
-    def refresh_segment(self, complete=False):
-
-        logger.debug("Refreshing segment:")
-        self.init_tokens()
-        self.last_attend_frame = -self.cfg.rewind_threshold       
-        # self.detected_language = None
-        self.cumulative_time_offset = 0.0
-        self.init_context()
-        logger.debug(f"Context: {self.context}")
-        if not complete and len(self.segments) > 2:
-            self.segments = self.segments[-2:]
-        else:
-            logger.debug("removing all segments.")
-            self.segments = []
-        self.log_segments += 1
-
-        self.pending_incomplete_tokens = []
-
-    def fire_at_boundary(self, chunked_encoder_feature: torch.Tensor):
-        if self.always_fire: return True
-        if self.never_fire: return False
-        return fire_at_boundary(chunked_encoder_feature, self.CIFLinear)
+    def init_context(self):
+        kw = {
+            'tokenizer': self.tokenizer,
+            'device': self.model.device,
+            'prefix_token_ids': [self.tokenizer.sot_prev],
+        }
+        self.state.context = TokenBuffer.empty(**kw)
+        if self.cfg.static_init_prompt is not None:
+            self.state.context = TokenBuffer.from_text(self.cfg.static_init_prompt, **kw)
+        if self.cfg.init_prompt is not None:
+            self.state.context.text += self.cfg.init_prompt
 
+    def insert_audio(self, segment=None):
+        if segment is not None:
+            self.state.segments.append(segment)
+        removed_len = 0
+        segments_len = self.segments_len()
+        while len(self.state.segments) > 1 and segments_len > self.cfg.audio_max_len:
+            removed_len = self.state.segments[0].shape[0] / 16000
+            segments_len -= removed_len
+            self.state.last_attend_frame -= int(TOKENS_PER_SECOND * removed_len)
+            self.state.cumulative_time_offset += removed_len
+            self.state.segments = self.state.segments[1:]
+            logger.debug(
+                f"remove segments: {len(self.state.segments)} {len(self.state.tokens)}, "
+                f"cumulative offset: {self.state.cumulative_time_offset:.2f}s"
+            )
+            if len(self.state.tokens) > 1:
+                self.state.context.append_token_ids(self.state.tokens[1][0, :].tolist())
+                self.state.tokens = [self.state.initial_tokens] + self.state.tokens[2:]
+        return removed_len
 
     def _current_tokens(self):
-
-        toks = self.tokens
-        # very first infer: duplicate start of seq to beam_size
+        toks = self.state.tokens
         if toks[0].shape[0] == 1:
-            toks[0] = toks[0].repeat_interleave(self.cfg.beam_size,dim=0)
-
-        if not self.context.is_empty():
-            context_toks = self.context.as_tensor_beam(self.cfg.beam_size, device=self.model.device)
+            toks[0] = toks[0].repeat_interleave(self.cfg.beam_size, dim=0)
+        if not self.state.context.is_empty():
+            context_toks = self.state.context.as_tensor_beam(
+                self.cfg.beam_size, device=self.model.device,
+            )
             toks = [context_toks] + toks
-
-        # make it one tensor
         if len(toks) > 1:
             current_tokens = torch.cat(toks, dim=1)
         else:
@@ -305,66 +201,19 @@ class AlignAtt:
         self.debug_print_tokens(current_tokens)
         return current_tokens
 
-
-    def debug_print_tokens(self, tokens):
-        for i in range(self.cfg.beam_size):
-            logger.debug(self.tokenizer.decode_with_timestamps(tokens[i].tolist()))
-
-    ### audio buffer 
-
-    def segments_len(self):
-        segments_len = sum(s.shape[0] for s in self.segments) / 16000
-        return segments_len
-
-    def _apply_minseglen(self):
-        segments_len = self.segments_len()
-        # wait for long enough audio to start
-        if segments_len < self.cfg.audio_min_len: 
-            logger.debug("waiting for next segment")
+    def fire_at_boundary(self, chunked_encoder_feature: torch.Tensor):
+        if self.state.always_fire:
+            return True
+        if self.state.never_fire:
             return False
-        return True
-
-    def insert_audio(self, segment=None):
-        if segment is not None:
-            self.segments.append(segment)
-
-        removed_len = 0
-        # len of audio is bigger than buffer_len. Going to remove the first segment
-        segments_len = self.segments_len()
-        while len(self.segments) > 1 and segments_len > self.cfg.audio_max_len:
-            removed_len = self.segments[0].shape[0] / 16000
-            segments_len -= removed_len
-            self.last_attend_frame -= int(TOKENS_PER_SECOND*removed_len)
-            self.cumulative_time_offset += removed_len  # Track cumulative time removed
-            self.segments = self.segments[1:]
-            logger.debug(f"remove segments: {len(self.segments)} {len(self.tokens)}, cumulative offset: {self.cumulative_time_offset:.2f}s")
-            if len(self.tokens) > 1:
-                self.context.append_token_ids(self.tokens[1][0,:].tolist())
-                self.tokens = [self.initial_tokens] + self.tokens[2:]
-        return removed_len
-
-    def _clean_cache(self):
-        '''clean the cache that stores the attention matrices and kv_cache.
-        It must be called every time after generation with the model.'''
-        # cleaning cache
-        self.dec_attns = []
-        self.kv_cache = {}
-        if self.decoder_type == "beam":
-            self.inference.kv_cache = self.kv_cache
-            self.token_decoder.reset()
+        return fire_at_boundary(chunked_encoder_feature, self.state.CIFLinear)
 
     @torch.no_grad()
     def lang_id(self, encoder_features):
-        """Language detection from encoder features.
-        This code is trimmed and copy-pasted from whisper.decoding.detect_language .
-        """
-    
-        # forward pass using a single token, startoftranscript
         n_audio = encoder_features.shape[0]
-        x = torch.tensor([[self.tokenizer.sot]] * n_audio).to(self.model.device)  # [n_audio, 1]
+        x = torch.tensor([[self.tokenizer.sot]] * n_audio).to(self.model.device)
         logits = self.model.logits(x, encoder_features)[:, 0]
 
-        # collect detected languages; suppress all non-language tokens
         mask = torch.ones(logits.shape[-1], dtype=torch.bool)
         mask[list(self.tokenizer.all_language_tokens)] = False
         logits[:, mask] = -np.inf
@@ -373,46 +222,31 @@ class AlignAtt:
         language_probs = [
             {
                 c: language_token_probs[i, j].item()
-                for j, c in zip(self.tokenizer.all_language_tokens, self.tokenizer.all_language_codes)
+                for j, c in zip(
+                    self.tokenizer.all_language_tokens,
+                    self.tokenizer.all_language_codes,
+                )
             }
             for i in range(n_audio)
         ]
-
         single = encoder_features.ndim == 2
         if single:
             language_tokens = language_tokens[0]
             language_probs = language_probs[0]
-
         self._clean_cache()
         return language_tokens, language_probs
 
-    ### transcription / translation
+    def _concat_segments(self):
+        if len(self.state.segments) > 1:
+            return torch.cat(self.state.segments, dim=0)
+        return self.state.segments[0]
 
-    @torch.no_grad()
-    def infer(self, is_last=False):
-        new_segment = True
-        if len(self.segments) == 0:
-            logger.debug("No segments, nothing to do")
-            return []
-        if not self._apply_minseglen():
-            logger.debug(f"applied minseglen {self.cfg.audio_min_len} > {self.segments_len()}.")
-            input_segments = torch.cat(self.segments, dim=0)
-            return []
-
-        # input_segments is concatenation of audio, it's one array
-        if len(self.segments) > 1:
-            input_segments = torch.cat(self.segments, dim=0)
-        else:
-            input_segments = self.segments[0]
-
-        beg_encode = time()
+    def _encode(self, input_segments):
         if self.use_mlcore:
             coreml_encoder, coreml_input_name, coreml_output_name = self.coreml_encoder_tuple
             mel_padded = log_mel_spectrogram(
-                input_segments,
-                n_mels=self.model.dims.n_mels,
-                padding=N_SAMPLES,
-                device="cpu",
+                input_segments, n_mels=self.model.dims.n_mels,
+                padding=N_SAMPLES, device="cpu",
             ).unsqueeze(0)
             mel = pad_or_trim(mel_padded, N_FRAMES)
             content_mel_len = int((mel_padded.shape[2] - mel.shape[2]) / 2)
@@ -424,237 +258,151 @@ class AlignAtt:
             else:
                 encoder_feature_np = next(iter(coreml_outputs.values()))
             encoder_feature = torch.as_tensor(
-                np.array(encoder_feature_np),
-                device=self.device,
+                np.array(encoder_feature_np), device=self.device,
             )
         if self.mlx_encoder:
-            mlx_mel_padded = mlx_log_mel_spectrogram(audio=input_segments.detach(), n_mels=self.model.dims.n_mels, padding=N_SAMPLES)
+            mlx_mel_padded = mlx_log_mel_spectrogram(
+                audio=input_segments.detach(),
+                n_mels=self.model.dims.n_mels, padding=N_SAMPLES,
+            )
             mlx_mel = mlx_pad_or_trim(mlx_mel_padded, N_FRAMES, axis=-2)
             mlx_encoder_feature = self.mlx_encoder.encoder(mlx_mel[None])
             encoder_feature = torch.as_tensor(mlx_encoder_feature)
-            content_mel_len = int((mlx_mel_padded.shape[0] - mlx_mel.shape[0])/2)
+            content_mel_len = int((mlx_mel_padded.shape[0] - mlx_mel.shape[0]) / 2)
         elif self.fw_encoder:
-            audio_length_seconds = len(input_segments) / 16000   
-            content_mel_len = int(audio_length_seconds * 100)//2      
-            mel_padded_2 = self.fw_feature_extractor(waveform=input_segments.numpy(), padding=N_SAMPLES)[None, :]
+            audio_length_seconds = len(input_segments) / 16000
+            content_mel_len = int(audio_length_seconds * 100) // 2
+            mel_padded_2 = self.fw_feature_extractor(
+                waveform=input_segments.numpy(), padding=N_SAMPLES,
+            )[None, :]
             mel = fw_pad_or_trim(mel_padded_2, N_FRAMES, axis=-1)
             encoder_feature_ctranslate = self.fw_encoder.encode(mel)
-            if self.device == 'cpu': #it seems that on gpu, passing StorageView to torch.as_tensor fails and wrapping in the array works
+            if self.device == 'cpu':
                 encoder_feature_ctranslate = np.array(encoder_feature_ctranslate)
             try:
                 encoder_feature = torch.as_tensor(encoder_feature_ctranslate, device=self.device)
-            except TypeError: # Normally the cpu condition should prevent having exceptions, but just in case:
-                encoder_feature = torch.as_tensor(np.array(encoder_feature_ctranslate), device=self.device)
+            except TypeError:
+                # Some numpy/ctranslate2 versions produce object_ dtype arrays; force float32
+                arr = np.array(encoder_feature_ctranslate)
+                if arr.dtype == np.object_:
+                    arr = np.array(arr.tolist(), dtype=np.float32)
+                encoder_feature = torch.as_tensor(arr, device=self.device)
         else:
-            # mel + padding to 30s
-            mel_padded = log_mel_spectrogram(input_segments, n_mels=self.model.dims.n_mels, padding=N_SAMPLES, 
-                                                device=self.device).unsqueeze(0)
-            # trim to 3000
+            mel_padded = log_mel_spectrogram(
+                input_segments, n_mels=self.model.dims.n_mels,
+                padding=N_SAMPLES, device=self.device,
+            ).unsqueeze(0)
             mel = pad_or_trim(mel_padded, N_FRAMES)
-            # the len of actual audio
-            content_mel_len = int((mel_padded.shape[2] - mel.shape[2])/2)
+            content_mel_len = int((mel_padded.shape[2] - mel.shape[2]) / 2)
             encoder_feature = self.model.encoder(mel)
-        end_encode = time()
-        # print('Encoder duration:', end_encode-beg_encode)
-                
-        if self.cfg.language == "auto" and self.detected_language is None and self.first_timestamp:
-            seconds_since_start = self.segments_len() - self.first_timestamp
-            if seconds_since_start >= 2.0:
-                language_tokens, language_probs = self.lang_id(encoder_feature) 
-                top_lan, p = max(language_probs[0].items(), key=lambda x: x[1])
-                print(f"Detected language: {top_lan} with p={p:.4f}")
-                self.create_tokenizer(top_lan)
-                self.last_attend_frame = -self.cfg.rewind_threshold
-                self.cumulative_time_offset = 0.0
-                self.init_tokens()
-                self.init_context()
-                self.detected_language = top_lan
-                logger.info(f"Tokenizer language: {self.tokenizer.language}, {self.tokenizer.sot_sequence_including_notimestamps}")
+        return encoder_feature, content_mel_len
 
-        self.trim_context()
-        current_tokens = self._current_tokens()
-   
-        fire_detected = self.fire_at_boundary(encoder_feature[:, :content_mel_len, :])
+    def _init_sum_logprobs(self):
+        return torch.zeros(self.cfg.beam_size, device=self.device)
 
-
-        sum_logprobs = torch.zeros(self.cfg.beam_size, device=self.device)
-        completed = False
-        # punctuation_stop = False
-
-        attn_of_alignment_heads = None
-        most_attended_frame = None
-
-        token_len_before_decoding = current_tokens.shape[1]
-        
-        l_absolute_timestamps = []
-        
-        while not completed and current_tokens.shape[1] < self.max_text_len: # bos is 3 tokens
-
-            if new_segment:
-                tokens_for_logits = current_tokens
-            else:
-                # only need to use the last token except in the first forward pass
-                tokens_for_logits = current_tokens[:,-1:]
-
-            logits = self.logits(tokens_for_logits, encoder_feature) # B, len(tokens), token dict size
-
-            if new_segment and self.tokenizer.no_speech is not None:
-                probs_at_sot = logits[:, self.sot_index, :].float().softmax(dim=-1)
-                no_speech_probs = probs_at_sot[:, self.tokenizer.no_speech].tolist()
-                if no_speech_probs[0] > self.cfg.nonspeech_prob:
-                    logger.info("no speech, stop")
-                    break
-
-            logits = logits[:, -1, :] # logits for the last token
-
-            # supress blank tokens only at the beginning of the segment
-            if new_segment:
-                logits[:, self.tokenizer.encode(" ") + [self.tokenizer.eot]] = -np.inf
-            new_segment = False
-            self.suppress_tokens(logits)
-            current_tokens, completed = self.token_decoder.update(current_tokens, logits, sum_logprobs)
-
-            logger.debug(f"Decoding completed: {completed}, sum_logprobs: {sum_logprobs.tolist()}, tokens: ")
-            self.debug_print_tokens(current_tokens)
-
-            attn_of_alignment_heads = [[] for _ in range(self.num_align_heads)]
-            for i, attn_mat in enumerate(self.dec_attns):
-                layer_rank = int(i % len(self.model.decoder.blocks))
-                align_heads_in_layer = self.align_source.get(layer_rank, [])
-                if len(align_heads_in_layer) == 0:
-                    continue
-                for align_head_rank, head_id in align_heads_in_layer:
-                    if self.cfg.beam_size == 1:
-                        a = attn_mat[head_id, :, :]
-                        a = a.unsqueeze(0)
-                    else:
-                        a = attn_mat[:, head_id, :, :]
-                    attn_of_alignment_heads[align_head_rank].append(a)
-            tmp = []
-            for mat in attn_of_alignment_heads:
-                t = torch.cat(mat, dim=1)
-                tmp.append(t) 
-            attn_of_alignment_heads = torch.stack(tmp, dim=1)
-            std, mean = torch.std_mean(attn_of_alignment_heads, dim=-2, keepdim=True, unbiased=False)
-            attn_of_alignment_heads = (attn_of_alignment_heads - mean) / std
-            attn_of_alignment_heads = median_filter(attn_of_alignment_heads, 7) # from whisper.timing
-            attn_of_alignment_heads = attn_of_alignment_heads.mean(dim=1)
-            attn_of_alignment_heads = attn_of_alignment_heads[:,:, :content_mel_len]
-
-            # for each beam, the most attended frame is:
-            most_attended_frames = torch.argmax(attn_of_alignment_heads[:,-1,:], dim=-1)
-            
-            # Calculate absolute timestamps accounting for cumulative offset
-            absolute_timestamps = [(frame * 0.02 + self.cumulative_time_offset) for frame in most_attended_frames.tolist()]
-            
-            logger.debug(str(most_attended_frames.tolist()) + " most att frames")
-            logger.debug(f"Absolute timestamps: {absolute_timestamps} (offset: {self.cumulative_time_offset:.2f}s)")
-
-            most_attended_frame = most_attended_frames[0].item()
-            l_absolute_timestamps.append(absolute_timestamps[0])
-
-            logger.debug("current tokens" + str(current_tokens.shape))
-            if completed:
-            #    # stripping the last token, the eot
-                current_tokens = current_tokens[:, :-1]
-                break
-            
-            # for some rare cases where the attention fails
-            if not is_last and self.last_attend_frame - most_attended_frame > self.cfg.rewind_threshold:
-                # TODO: check this
-                if current_tokens.shape[1] > 1 and current_tokens[0, -2] >= DEC_PAD:
-                    logger.debug("ommit rewinding from special tokens")
-                    self.last_attend_frame = most_attended_frame
-                else:
-                    logger.debug(
-                        f"[rewind detected] current attention pos: {most_attended_frame}, "
-                        f"last attention pos: {self.last_attend_frame}; omit this segment")
-                    self.last_attend_frame = -self.cfg.rewind_threshold
-                    current_tokens = torch.cat(self.tokens, dim=1) if len(self.tokens) > 0 else self.tokens[0]
-                    break
-            else:
-                self.last_attend_frame = most_attended_frame
-
-            if content_mel_len - most_attended_frame <= (4 if is_last else self.cfg.frame_threshold):
-                logger.debug(f"attention reaches the end: {most_attended_frame}/{content_mel_len}")
-                # stripping the last token, the one that is attended too close to the end
-                current_tokens = current_tokens[:, :-1]
-                break
-        
-            # debug print
-            for i in range(self.cfg.beam_size):
-                logger.debug("attn: {}, current pos: {}, current token: {}({})".format(
-                    attn_of_alignment_heads.shape if attn_of_alignment_heads is not None else None,
-                    most_attended_frames[i], 
-                    current_tokens[i, -1].item(),
-                    self.tokenizer.decode([current_tokens[i, -1].item()])
-                ))
-
-        tokens_to_split = current_tokens[0, token_len_before_decoding:]
-
-        # Prepend pending tokens from previous chunk if any
-        if self.pending_incomplete_tokens:
-            logger.debug(f"[UTF-8 Fix] Prepending {len(self.pending_incomplete_tokens)} pending tokens: {self.pending_incomplete_tokens}")
-            pending_tensor = torch.tensor(self.pending_incomplete_tokens, dtype=torch.long, device=self.device)
-            tokens_to_split = torch.cat([pending_tensor, tokens_to_split])
-
-        if fire_detected or is_last: #or punctuation_stop:
-            new_hypothesis = tokens_to_split.flatten().tolist()
-            split_words, split_tokens = self.tokenizer.split_to_word_tokens(new_hypothesis)
-        else:
-            # going to truncate the tokens after the last space
-            split_words, split_tokens = self.tokenizer.split_to_word_tokens(tokens_to_split.tolist())
-            if len(split_words) > 1:
-                new_hypothesis = [i for sublist in split_tokens[:-1] for i in sublist]  
-            else:
-                new_hypothesis = []
-
-
-        logger.debug(f"new_hypothesis: {new_hypothesis}")
-        new_tokens = torch.tensor([new_hypothesis], dtype=torch.long).repeat_interleave(self.cfg.beam_size, dim=0).to(
-            device=self.device,
-        )
-        self.tokens.append(new_tokens)
-
-        logger.info(f"Output: {self.tokenizer.decode(new_hypothesis)}")
-        
-        self._clean_cache()
-
-        if len(l_absolute_timestamps) >=2 and self.first_timestamp is None:
-            self.first_timestamp = l_absolute_timestamps[0]
-            
-
-        timestamped_words = []
-        timestamp_idx = 0
-        replacement_char = "\ufffd"
-        for word, word_tokens in zip(split_words, split_tokens):
-            # Skip words containing incomplete UTF-8 from client output
-            if replacement_char in word:
-                logger.warning(f"[UTF-8 Filter] Skipping incomplete word from client output: {repr(word)}")
-                timestamp_idx += len(word_tokens)
-                continue
-
-            try:
-                current_timestamp = l_absolute_timestamps[timestamp_idx]
-            except:
-                pass
-            timestamp_idx += len(word_tokens)
-
-            timestamp_entry = ASRToken(
-                    start=round(current_timestamp, 2),
-                    end=round(current_timestamp + 0.1, 2),
-                    text= word,
-                    speaker=self.speaker,
-                    detected_language=self.detected_language
-                ).with_offset(
-                    self.global_time_offset
+    def _get_logits_and_cross_attn(self, tokens, encoder_feature):
+        if self.state.decoder_type == "greedy":
+            return self.model.decoder(
+                tokens, encoder_feature,
+                kv_cache=self.state.kv_cache,
+                return_cross_attn=True,
+            )
+        else:
+            logger.debug(f"Logits shape: {tokens.shape}")
+            return self.state.inference.logits(
+                tokens, encoder_feature, return_cross_attn=True,
             )
-            timestamped_words.append(timestamp_entry)
 
-        # Hold incomplete tokens for next chunk
-        self.pending_incomplete_tokens = []
-        if split_words and replacement_char in split_words[-1]:
-            self.pending_incomplete_tokens = split_tokens[-1]
-            logger.warning(f"[UTF-8 Fix] Holding {len(self.pending_incomplete_tokens)} incomplete tokens for next chunk: {self.pending_incomplete_tokens}")
+    def _check_no_speech(self, logits):
+        if self.tokenizer.no_speech is not None:
+            probs_at_sot = logits[:, self.state.sot_index, :].float().softmax(dim=-1)
+            no_speech_probs = probs_at_sot[:, self.tokenizer.no_speech].tolist()
+            if no_speech_probs[0] > self.cfg.nonspeech_prob:
+                logger.info("no speech, stop")
+                return True
+        return False
 
-        return timestamped_words
+    def _suppress_blank_tokens(self, logits):
+        logits[:, self.tokenizer.encode(" ") + [self.tokenizer.eot]] = -np.inf
+        return logits
+
+    def _apply_token_suppression(self, logits):
+        self.state.suppress_tokens_fn(logits)
+        return logits
+
+    def _update_tokens(self, current_tokens, logits, sum_logprobs):
+        return self.state.token_decoder.update(current_tokens, logits, sum_logprobs)
+
+    def _process_cross_attention(
+        self, cross_attns: List, content_mel_len: int,
+    ) -> torch.Tensor:
+        attn_of_alignment_heads = [[] for _ in range(self.state.num_align_heads)]
+        num_decoder_layers = len(self.model.decoder.blocks)
+
+        if cross_attns and isinstance(cross_attns[0], list):
+            flattened_attns = [attn for layer_list in cross_attns for attn in layer_list]
+        else:
+            flattened_attns = cross_attns
+
+        for idx, attn_mat in enumerate(flattened_attns):
+            layer_rank = idx % num_decoder_layers
+            align_heads_in_layer = self.state.align_source.get(layer_rank, [])
+            if not align_heads_in_layer:
+                continue
+            attn_mat = F.softmax(attn_mat, dim=-1)
+            for align_head_rank, head_id in align_heads_in_layer:
+                if self.cfg.beam_size == 1:
+                    if attn_mat.dim() == 4:
+                        a = attn_mat[0, head_id, :, :]
+                    else:
+                        a = attn_mat[head_id, :, :]
+                    a = a.unsqueeze(0)
+                else:
+                    a = attn_mat[:, head_id, :, :]
+                attn_of_alignment_heads[align_head_rank].append(a)
+
+        tmp = []
+        for mat in attn_of_alignment_heads:
+            if mat:
+                tmp.append(torch.cat(mat, dim=1))
+        if not tmp:
+            return torch.zeros(self.cfg.beam_size, 1, content_mel_len, device=self.device)
+
+        attn_of_alignment_heads = torch.stack(tmp, dim=1)
+        std, mean = torch.std_mean(
+            attn_of_alignment_heads, dim=-2, keepdim=True, unbiased=False,
+        )
+        attn_of_alignment_heads = (attn_of_alignment_heads - mean) / (std + 1e-8)
+        attn_of_alignment_heads = median_filter(attn_of_alignment_heads, 7)
+        attn_of_alignment_heads = attn_of_alignment_heads.mean(dim=1)
+        attn_of_alignment_heads = attn_of_alignment_heads[:, :, :content_mel_len]
+        return attn_of_alignment_heads
+
+    def _get_attended_frames(self, attn):
+        most_attended_frames = torch.argmax(attn[:, -1, :], dim=-1)
+        return most_attended_frames.tolist(), most_attended_frames[0].item()
+
+    def _is_special_token(self, current_tokens):
+        return current_tokens[0, -2].item() >= DEC_PAD
+
+    def _rewind_tokens(self):
+        if len(self.state.tokens) > 0:
+            return torch.cat(self.state.tokens, dim=1)
+        return self.state.tokens[0]
+
+    def _tokens_to_list(self, current_tokens, start_col):
+        return current_tokens[0, start_col:].flatten().tolist()
+
+    def _make_new_tokens_tensor(self, hypothesis):
+        return (
+            torch.tensor([hypothesis], dtype=torch.long)
+            .repeat_interleave(self.cfg.beam_size, dim=0)
+            .to(device=self.device)
+        )
+
+    def _evaluate(self, tensor):
+        pass  # No-op for PyTorch
+
+    @torch.no_grad()
+    def infer(self, is_last=False):
+        return super().infer(is_last)

whisperlivekit/simul_whisper/token_buffer.py

@@ -1,5 +1,8 @@
-import torch
 import sys
+
+import torch
+
+
 class TokenBuffer:
 
     def __init__(self, text="", tokenizer=None, device=None, prefix_token_ids=[]):

whisperlivekit/thread_safety.py

@@ -0,0 +1,139 @@
+"""
+Thread Safety Configuration for WhisperLiveKit
+
+This module provides thread safety configuration and utilities.
+
+Environment Variables:
+    WHISPERLIVEKIT_MODEL_LOCK: Enable/disable model locking (default: 1)
+        Set to "0" to disable for single-connection deployments
+
+    WHISPERLIVEKIT_LOCK_TIMEOUT: Lock acquisition timeout in seconds (default: 30)
+
+Usage:
+    # Enable model locking (default)
+    export WHISPERLIVEKIT_MODEL_LOCK=1
+
+    # Disable for single-connection deployment
+    export WHISPERLIVEKIT_MODEL_LOCK=0
+
+    # Custom timeout
+    export WHISPERLIVEKIT_LOCK_TIMEOUT=60
+"""
+
+import os
+import logging
+import threading
+
+logger = logging.getLogger(__name__)
+
+# Configuration
+USE_MODEL_LOCK = os.environ.get("WHISPERLIVEKIT_MODEL_LOCK", "1") == "1"
+LOCK_TIMEOUT = float(os.environ.get("WHISPERLIVEKIT_LOCK_TIMEOUT", "30.0"))
+
+# Global model lock
+_model_lock = threading.Lock()
+
+# Log configuration on import
+if USE_MODEL_LOCK:
+    logger.info(f"Model locking ENABLED (timeout: {LOCK_TIMEOUT}s)")
+    logger.info("For single-connection deployments, set WHISPERLIVEKIT_MODEL_LOCK=0")
+else:
+    logger.warning("Model locking DISABLED - only safe for single-connection deployments")
+
+
+def get_model_lock():
+    """Get the global model lock instance"""
+    return _model_lock
+
+
+def acquire_model_lock(timeout=None):
+    """
+    Acquire model lock with timeout.
+
+    Args:
+        timeout: Lock acquisition timeout (default: use LOCK_TIMEOUT)
+
+    Returns:
+        bool: True if lock acquired, False on timeout
+    """
+    if not USE_MODEL_LOCK:
+        return True
+
+    timeout = timeout or LOCK_TIMEOUT
+    acquired = _model_lock.acquire(timeout=timeout)
+
+    if not acquired:
+        logger.error(f"Failed to acquire model lock within {timeout}s")
+
+    return acquired
+
+
+def release_model_lock():
+    """Release model lock"""
+    if not USE_MODEL_LOCK:
+        return
+
+    try:
+        _model_lock.release()
+    except RuntimeError:
+        # Lock not held - this is fine
+        pass
+
+
+class ModelLockContext:
+    """Context manager for model lock"""
+
+    def __init__(self, timeout=None):
+        self.timeout = timeout
+        self.acquired = False
+
+    def __enter__(self):
+        self.acquired = acquire_model_lock(self.timeout)
+        return self.acquired
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if self.acquired:
+            release_model_lock()
+        return False
+
+
+# Concurrency recommendations
+RECOMMENDED_CONNECTIONS_PER_WORKER = 1 if USE_MODEL_LOCK else 1
+RECOMMENDED_WORKERS = 4
+
+def print_deployment_recommendations():
+    """Print recommended deployment configuration"""
+    print("\n" + "="*60)
+    print("WhisperLiveKit Deployment Recommendations")
+    print("="*60)
+
+    if USE_MODEL_LOCK:
+        print("⚠️  Model locking is ENABLED")
+        print("   This serializes inference across connections.")
+        print()
+        print("Recommended deployment:")
+        print(f"  gunicorn -w {RECOMMENDED_WORKERS} \\")
+        print("    -k uvicorn.workers.UvicornWorker \\")
+        print("    --worker-connections 1 \\")
+        print("    whisperlivekit.basic_server:app")
+        print()
+        print("Expected capacity:")
+        print(f"  - {RECOMMENDED_WORKERS} concurrent users (1 per worker)")
+        print(f"  - Memory: ~{RECOMMENDED_WORKERS}x model size")
+    else:
+        print("✅ Model locking is DISABLED")
+        print("   ⚠️  ONLY safe for single-connection deployments")
+        print()
+        print("Recommended deployment:")
+        print("  uvicorn whisperlivekit.basic_server:app \\")
+        print("    --host 0.0.0.0 --port 8000 \\")
+        print("    --workers 1")
+        print()
+        print("Expected capacity:")
+        print("  - 1 concurrent user only")
+
+    print("="*60 + "\n")
+
+
+if __name__ == "__main__":
+    print_deployment_recommendations()

whisperlivekit/timed_objects.py

@@ -1,6 +1,6 @@
 from dataclasses import dataclass, field
-from typing import Optional, List, Union, Dict, Any
 from datetime import timedelta
+from typing import Any, Dict, List, Optional, Union
 
 PUNCTUATION_MARKS = {'.', '!', '?', '。', '！', '？'}
 
@@ -39,10 +39,11 @@ class TimedText(Timed):
 
 @dataclass()
 class ASRToken(TimedText):
-    
+    probability: Optional[float] = None
+
     def with_offset(self, offset: float) -> "ASRToken":
         """Return a new token with the time offset added."""
-        return ASRToken(self.start + offset, self.end + offset, self.text, self.speaker, detected_language=self.detected_language)
+        return ASRToken(self.start + offset, self.end + offset, self.text, self.speaker, detected_language=self.detected_language, probability=self.probability)
 
     def is_silence(self) -> bool:
         return False
@@ -114,6 +115,9 @@ class Segment(TimedText):
     end: Optional[float]
     text: Optional[str]
     speaker: Optional[str]
+    tokens: Optional[ASRToken] = None
+    translation: Optional[Translation] = None
+
     @classmethod
     def from_tokens(
         cls,
@@ -141,17 +145,13 @@ class Segment(TimedText):
                 speaker=-1,
                 detected_language=start_token.detected_language
             )
+
     def is_silence(self) -> bool:
         """True when this segment represents a silence gap."""
         return self.speaker == -2
 
-
-@dataclass
-class Line(TimedText):
-    translation: str = ''
-    
     def to_dict(self) -> Dict[str, Any]:
-        """Serialize the line for frontend consumption."""
+        """Serialize the segment for frontend consumption."""
         _dict: Dict[str, Any] = {
             'speaker': int(self.speaker) if self.speaker != -1 else 1,
             'text': self.text,
@@ -163,29 +163,13 @@ class Line(TimedText):
         if self.detected_language:
             _dict['detected_language'] = self.detected_language
         return _dict
-    
-    def build_from_tokens(self, tokens: List[ASRToken]) -> "Line":
-        """Populate line attributes from a contiguous token list."""
-        self.text = ''.join([token.text for token in tokens])
-        self.start = tokens[0].start
-        self.end = tokens[-1].end
-        self.speaker = 1
-        self.detected_language = tokens[0].detected_language
-        return self
 
-    def build_from_segment(self, segment: Segment) -> "Line":
-        """Populate the line fields from a pre-built segment."""
-        self.text = segment.text
-        self.start = segment.start
-        self.end = segment.end
-        self.speaker = segment.speaker
-        self.detected_language = segment.detected_language
-        return self
 
-    def is_silent(self) -> bool:
-        return self.speaker == -2
+@dataclass
+class PuncSegment(Segment):
+    pass
 
-class SilentLine(Line):
+class SilentSegment(Segment):
     def __init__(self, *args: Any, **kwargs: Any) -> None:
         super().__init__(*args, **kwargs)
         self.speaker = -2
@@ -196,7 +180,7 @@ class SilentLine(Line):
 class FrontData():
     status: str = ''
     error: str = ''
-    lines: list[Line] = field(default_factory=list)
+    lines: list[Segment] = field(default_factory=list)
     buffer_transcription: str = ''
     buffer_diarization: str = ''
     buffer_translation: str = ''

whisperlivekit/tokens_alignment.py

@@ -1,7 +1,9 @@
 from time import time
-from typing import Optional, List, Tuple, Union, Any
+from typing import Any, List, Optional, Tuple, Union
 
-from whisperlivekit.timed_objects import Line, SilentLine, ASRToken, SpeakerSegment, Silence, TimedText, Segment
+from whisperlivekit.timed_objects import (ASRToken, Segment, PuncSegment, Silence,
+                                          SilentSegment, SpeakerSegment,
+                                          TimedText)
 
 
 class TokensAlignment:
@@ -25,6 +27,14 @@ class TokensAlignment:
         self.sep: str = sep if sep is not None else ' '
         self.beg_loop: Optional[float] = None
 
+        self.validated_segments: List[Segment] = []
+        self.current_line_tokens: List[ASRToken] = []
+        self.diarization_buffer: List[ASRToken] = []
+
+        self.last_punctuation = None
+        self.last_uncompleted_punc_segment: PuncSegment = None
+        self.unvalidated_tokens: PuncSegment = []
+
     def update(self) -> None:
         """Drain state buffers into the running alignment context."""
         self.new_tokens, self.state.new_tokens = self.state.new_tokens, []
@@ -37,27 +47,30 @@ class TokensAlignment:
         self.all_translation_segments.extend(self.new_translation)
         self.new_translation_buffer = self.state.new_translation_buffer
 
-    def add_translation(self, line: Line) -> None:
-        """Append translated text segments that overlap with a line."""
+    def add_translation(self, segment: Segment) -> None:
+        """Append translated text segments that overlap with a segment."""
+        if segment.translation is None:
+            segment.translation = ''
         for ts in self.all_translation_segments:
-            if ts.is_within(line):
-                line.translation += ts.text + (self.sep if ts.text else '')
-            elif line.translation:
+            if ts.is_within(segment):
+                if ts.text:
+                    segment.translation += ts.text + self.sep
+            elif segment.translation:
                 break
 
 
-    def compute_punctuations_segments(self, tokens: Optional[List[ASRToken]] = None) -> List[Segment]:
+    def compute_punctuations_segments(self, tokens: Optional[List[ASRToken]] = None) -> List[PuncSegment]:
         """Group tokens into segments split by punctuation and explicit silence."""
         segments = []
         segment_start_idx = 0
         for i, token in enumerate(self.all_tokens):
             if token.is_silence():
-                previous_segment = Segment.from_tokens(
+                previous_segment = PuncSegment.from_tokens(
                         tokens=self.all_tokens[segment_start_idx: i],
                     )
                 if previous_segment:
                     segments.append(previous_segment)
-                segment = Segment.from_tokens(
+                segment = PuncSegment.from_tokens(
                     tokens=[token],
                     is_silence=True
                 )
@@ -65,19 +78,47 @@ class TokensAlignment:
                 segment_start_idx = i+1
             else:
                 if token.has_punctuation():
-                    segment = Segment.from_tokens(
+                    segment = PuncSegment.from_tokens(
                         tokens=self.all_tokens[segment_start_idx: i+1],
                     )
                     segments.append(segment)
                     segment_start_idx = i+1
 
-        final_segment = Segment.from_tokens(
+        final_segment = PuncSegment.from_tokens(
             tokens=self.all_tokens[segment_start_idx:],
         )
         if final_segment:
             segments.append(final_segment)
         return segments
 
+    def compute_new_punctuations_segments(self) -> List[PuncSegment]:
+        new_punc_segments = []
+        segment_start_idx = 0
+        self.unvalidated_tokens += self.new_tokens
+        for i, token in enumerate(self.unvalidated_tokens):
+            if token.is_silence():
+                previous_segment = PuncSegment.from_tokens(
+                        tokens=self.unvalidated_tokens[segment_start_idx: i],
+                    )
+                if previous_segment:
+                    new_punc_segments.append(previous_segment)
+                segment = PuncSegment.from_tokens(
+                    tokens=[token],
+                    is_silence=True
+                )
+                new_punc_segments.append(segment)
+                segment_start_idx = i+1
+            else:
+                if token.has_punctuation():
+                    segment = PuncSegment.from_tokens(
+                        tokens=self.unvalidated_tokens[segment_start_idx: i+1],
+                    )
+                    new_punc_segments.append(segment)
+                    segment_start_idx = i+1
+
+        self.unvalidated_tokens = self.unvalidated_tokens[segment_start_idx:]
+        return new_punc_segments
+
 
     def concatenate_diar_segments(self) -> List[SpeakerSegment]:
         """Merge consecutive diarization slices that share the same speaker."""
@@ -100,8 +141,8 @@ class TokensAlignment:
 
         return max(0, end - start)
 
-    def get_lines_diarization(self) -> Tuple[List[Line], str]:
-        """Build lines when diarization is enabled and track overflow buffer."""
+    def get_lines_diarization(self) -> Tuple[List[Segment], str]:
+        """Build segments when diarization is enabled and track overflow buffer."""
         diarization_buffer = ''
         punctuation_segments = self.compute_punctuations_segments()
         diarization_segments = self.concatenate_diar_segments()
@@ -119,18 +160,18 @@ class TokensAlignment:
                             max_overlap_speaker = diarization_segment.speaker + 1
                     punctuation_segment.speaker = max_overlap_speaker
         
-        lines = []
+        segments = []
         if punctuation_segments:
-            lines = [Line().build_from_segment(punctuation_segments[0])]
+            segments = [punctuation_segments[0]]
             for segment in punctuation_segments[1:]:
-                if segment.speaker == lines[-1].speaker:
-                    if lines[-1].text:
-                        lines[-1].text += segment.text
-                    lines[-1].end = segment.end
+                if segment.speaker == segments[-1].speaker:
+                    if segments[-1].text:
+                        segments[-1].text += segment.text
+                    segments[-1].end = segment.end
                 else:
-                    lines.append(Line().build_from_segment(segment))
+                    segments.append(segment)
 
-        return lines, diarization_buffer
+        return segments, diarization_buffer
 
 
     def get_lines(
@@ -138,40 +179,42 @@ class TokensAlignment:
             diarization: bool = False,
             translation: bool = False,
             current_silence: Optional[Silence] = None
-        ) -> Tuple[List[Line], str, Union[str, TimedText]]:
-        """Return the formatted lines plus buffers, optionally with diarization/translation."""
+        ) -> Tuple[List[Segment], str, Union[str, TimedText]]:
+        """Return the formatted segments plus buffers, optionally with diarization/translation."""
         if diarization:
-            lines, diarization_buffer = self.get_lines_diarization()
+            segments, diarization_buffer = self.get_lines_diarization()
         else:
             diarization_buffer = ''
-            lines = []
-            current_line_tokens = []
-            for token in self.all_tokens:
-                if token.is_silence():
-                    if current_line_tokens:
-                        lines.append(Line().build_from_tokens(current_line_tokens))
-                        current_line_tokens = []
+            for token in self.new_tokens:
+                if isinstance(token, Silence):
+                    if self.current_line_tokens:
+                        self.validated_segments.append(Segment.from_tokens(self.current_line_tokens))
+                        self.current_line_tokens = []
+
                     end_silence = token.end if token.has_ended else time() - self.beg_loop
-                    if lines and lines[-1].is_silent():
-                        lines[-1].end = end_silence
+                    if self.validated_segments and self.validated_segments[-1].is_silence():
+                        self.validated_segments[-1].end = end_silence
                     else:
-                        lines.append(SilentLine(
-                            start = token.start,
-                            end = end_silence
+                        self.validated_segments.append(SilentSegment(
+                            start=token.start,
+                            end=end_silence
                         ))
                 else:
-                    current_line_tokens.append(token)
-            if current_line_tokens:
-                lines.append(Line().build_from_tokens(current_line_tokens))
+                    self.current_line_tokens.append(token)
+            
+            segments = list(self.validated_segments)
+            if self.current_line_tokens:
+                segments.append(Segment.from_tokens(self.current_line_tokens))
+
         if current_silence:
             end_silence = current_silence.end if current_silence.has_ended else time() - self.beg_loop
-            if lines and lines[-1].is_silent():
-                lines[-1].end = end_silence
+            if segments and segments[-1].is_silence():
+                segments[-1] = SilentSegment(start=segments[-1].start, end=end_silence)
             else:
-                lines.append(SilentLine(
-                    start = current_silence.start,
-                    end = end_silence
+                segments.append(SilentSegment(
+                    start=current_silence.start,
+                    end=end_silence
                 ))
         if translation:
-            [self.add_translation(line) for line in lines if not type(line) == Silence]
-        return lines, diarization_buffer, self.new_translation_buffer.text
+            [self.add_translation(segment) for segment in segments if not segment.is_silence()]
+        return segments, diarization_buffer, self.new_translation_buffer.text

whisperlivekit/voxtral_hf_streaming.py

@@ -0,0 +1,395 @@
+"""
+Voxtral Mini Realtime streaming backend using HuggingFace Transformers.
+
+Uses VoxtralRealtimeForConditionalGeneration with a background generate thread
+and queue-based audio feeding for real-time streaming transcription.
+Supports CUDA, CPU, and MPS devices.
+"""
+
+import logging
+import queue
+import sys
+import threading
+import time
+from typing import List, Optional, Tuple
+
+import numpy as np
+
+from whisperlivekit.timed_objects import ASRToken, Transcript
+
+logger = logging.getLogger(__name__)
+
+
+class VoxtralHFStreamingASR:
+    """Voxtral model holder using HuggingFace Transformers."""
+
+    sep = " "
+
+    def __init__(self, logfile=sys.stderr, **kwargs):
+        import torch
+        from transformers import (
+            AutoProcessor,
+            VoxtralRealtimeForConditionalGeneration,
+        )
+
+        self.logfile = logfile
+        self.transcribe_kargs = {}
+
+        lan = kwargs.get("lan", "auto")
+        self.original_language = None if lan == "auto" else lan
+
+        DEFAULT_MODEL = "mistralai/Voxtral-Mini-4B-Realtime-2602"
+        model_path = kwargs.get("model_dir") or kwargs.get("model_path")
+        if not model_path:
+            model_size = kwargs.get("model_size", "")
+            if model_size and ("/" in model_size or model_size.startswith(".")):
+                model_path = model_size
+            else:
+                model_path = DEFAULT_MODEL
+
+        t = time.time()
+        logger.info(f"Loading Voxtral model '{model_path}' via HF Transformers...")
+        self.processor = AutoProcessor.from_pretrained(model_path)
+        self.model = VoxtralRealtimeForConditionalGeneration.from_pretrained(
+            model_path,
+            torch_dtype=torch.bfloat16,
+            device_map="auto",
+        )
+        logger.info(f"Voxtral HF model loaded in {time.time() - t:.2f}s on {self.model.device}")
+
+        self.backend_choice = "voxtral"
+        self.tokenizer = None  # sentence tokenizer — not needed for streaming
+
+    def transcribe(self, audio):
+        pass
+
+
+class VoxtralHFStreamingOnlineProcessor:
+    """
+    Online processor for Voxtral streaming ASR via HuggingFace Transformers.
+
+    Uses a background thread running model.generate() with a queue-based
+    input_features_generator and TextIteratorStreamer for real-time output.
+    Each decoded token corresponds to ~80ms of audio.
+    """
+
+    SAMPLING_RATE = 16000
+
+    def __init__(self, asr: VoxtralHFStreamingASR, logfile=sys.stderr):
+        self.asr = asr
+        self.logfile = logfile
+        self.end = 0.0
+        self.buffer = []
+        self.audio_buffer = np.array([], dtype=np.float32)
+
+        processor = asr.processor
+        self._first_chunk_samples = processor.num_samples_first_audio_chunk
+        self._chunk_samples = processor.num_samples_per_audio_chunk
+        self._chunk_step = processor.raw_audio_length_per_tok
+        n_right_pad = processor.num_right_pad_tokens
+        if callable(n_right_pad):
+            n_right_pad = n_right_pad()
+        self._right_pad_samples = int(n_right_pad * processor.raw_audio_length_per_tok)
+        self._seconds_per_token = processor.raw_audio_length_per_tok / self.SAMPLING_RATE
+
+        self._reset_state()
+
+        logger.info(
+            f"[voxtral-hf] Initialized. first_chunk={self._first_chunk_samples} samples, "
+            f"chunk={self._chunk_samples}, step={self._chunk_step}, "
+            f"right_pad={self._right_pad_samples}"
+        )
+
+    def _reset_state(self):
+        self._pending_audio = np.zeros(0, dtype=np.float32)
+        self._audio_queue: queue.Queue = queue.Queue()
+        self._streamer_texts: List[str] = []
+        self._generate_thread: Optional[threading.Thread] = None
+        self._generate_started = False
+        self._generate_finished = False
+        self._generate_error: Optional[Exception] = None
+
+        # Text accumulation and word extraction
+        self._accumulated_text = ""
+        self._n_text_tokens_received = 0
+        self._n_committed_words = 0
+        self._global_time_offset = 0.0
+
+        # Lock for text state accessed from both generate thread and main thread
+        self._text_lock = threading.Lock()
+
+    # ── Interface methods ──
+
+    def insert_audio_chunk(self, audio: np.ndarray, audio_stream_end_time: float):
+        self.end = audio_stream_end_time
+        self._pending_audio = np.append(self._pending_audio, audio)
+        self.audio_buffer = self._pending_audio
+
+    def process_iter(self, is_last=False) -> Tuple[List[ASRToken], float]:
+        try:
+            return self._process_iter_inner(is_last)
+        except Exception as e:
+            logger.warning(f"[voxtral-hf] process_iter exception: {e}", exc_info=True)
+            return [], self.end
+
+    def get_buffer(self) -> Transcript:
+        """Return all uncommitted text as buffer."""
+        with self._text_lock:
+            text = self._accumulated_text
+        if not text:
+            return Transcript(start=None, end=None, text="")
+
+        words = text.split()
+        uncommitted = words[self._n_committed_words:]
+        if uncommitted:
+            return Transcript(start=self.end, end=self.end, text=" ".join(uncommitted))
+        return Transcript(start=None, end=None, text="")
+
+    def start_silence(self) -> Tuple[List[ASRToken], float]:
+        """Flush all uncommitted words when silence starts."""
+        self._drain_streamer()
+        words = self._flush_all_pending_words()
+        logger.info(f"[voxtral-hf] start_silence: flushed {len(words)} words")
+        return words, self.end
+
+    def end_silence(self, silence_duration: float, offset: float):
+        self._global_time_offset += silence_duration
+        self.end += silence_duration
+
+    def new_speaker(self, change_speaker):
+        self.start_silence()
+
+    def warmup(self, audio, init_prompt=""):
+        pass
+
+    def finish(self) -> Tuple[List[ASRToken], float]:
+        """Flush remaining audio with right-padding and stop the generate thread."""
+        # Add right-padding so the model can finish decoding
+        if self._right_pad_samples > 0:
+            right_pad = np.zeros(self._right_pad_samples, dtype=np.float32)
+            self._pending_audio = np.append(self._pending_audio, right_pad)
+
+        # Feed remaining audio
+        if self._generate_started and not self._generate_finished:
+            self._feed_pending_audio()
+            # Signal end of audio
+            self._audio_queue.put(None)
+            # Wait for generate to finish
+            if self._generate_thread is not None:
+                self._generate_thread.join(timeout=30.0)
+        elif not self._generate_started and len(self._pending_audio) >= self._first_chunk_samples:
+            # Never started but have enough audio — start and immediately finish
+            self._start_generate_thread()
+            self._feed_pending_audio()
+            self._audio_queue.put(None)
+            if self._generate_thread is not None:
+                self._generate_thread.join(timeout=30.0)
+
+        self._drain_streamer()
+        words = self._flush_all_pending_words()
+        logger.info(f"[voxtral-hf] finish: flushed {len(words)} words")
+        return words, self.end
+
+    # ── Generate thread management ──
+
+    def _start_generate_thread(self):
+        """Start model.generate() in a background thread with streaming."""
+        import torch
+        from transformers import TextIteratorStreamer
+
+        processor = self.asr.processor
+        model = self.asr.model
+
+        # Extract first chunk
+        first_chunk_audio = self._pending_audio[:self._first_chunk_samples]
+        self._pending_audio = self._pending_audio[self._first_chunk_samples:]
+
+        first_inputs = processor(
+            first_chunk_audio,
+            is_streaming=True,
+            is_first_audio_chunk=True,
+            return_tensors="pt",
+        )
+        first_inputs = first_inputs.to(model.device, dtype=model.dtype)
+
+        streamer = TextIteratorStreamer(
+            processor.tokenizer,
+            skip_prompt=True,
+            skip_special_tokens=True,
+        )
+        self._streamer = streamer
+
+        audio_queue = self._audio_queue
+
+        def input_features_gen():
+            yield first_inputs.input_features
+            while True:
+                chunk_audio = audio_queue.get()
+                if chunk_audio is None:
+                    break
+                inputs = processor(
+                    chunk_audio,
+                    is_streaming=True,
+                    is_first_audio_chunk=False,
+                    return_tensors="pt",
+                )
+                inputs = inputs.to(model.device, dtype=model.dtype)
+                yield inputs.input_features
+
+        def run_generate():
+            try:
+                with torch.no_grad():
+                    # Pass generator as input_features — the model detects GeneratorType
+                    # and internally converts it to input_features_generator
+                    generate_kwargs = {
+                        k: v for k, v in first_inputs.items()
+                        if k != "input_features"
+                    }
+                    model.generate(
+                        input_features=input_features_gen(),
+                        streamer=streamer,
+                        **generate_kwargs,
+                    )
+            except Exception as e:
+                logger.error(f"[voxtral-hf] generate error: {e}", exc_info=True)
+                self._generate_error = e
+            finally:
+                self._generate_finished = True
+
+        self._generate_thread = threading.Thread(target=run_generate, daemon=True)
+        self._generate_thread.start()
+        self._generate_started = True
+        logger.info("[voxtral-hf] generate thread started")
+
+    def _feed_pending_audio(self):
+        """Convert pending audio into properly-sized chunks for the generator."""
+        chunk_size = self._chunk_samples
+        step_size = self._chunk_step
+
+        while len(self._pending_audio) >= chunk_size:
+            chunk = self._pending_audio[:chunk_size]
+            self._audio_queue.put(chunk)
+            self._pending_audio = self._pending_audio[step_size:]
+
+        self.audio_buffer = self._pending_audio
+
+    def _drain_streamer(self):
+        """Non-blocking drain of all available text from the streamer."""
+        if not self._generate_started:
+            return
+
+        text_queue = self._streamer.text_queue
+        while True:
+            try:
+                text_fragment = text_queue.get_nowait()
+            except queue.Empty:
+                break
+            # TextIteratorStreamer uses None as end-of-stream sentinel
+            if text_fragment is None:
+                self._generate_finished = True
+                break
+            if text_fragment:
+                with self._text_lock:
+                    self._accumulated_text += text_fragment
+                self._n_text_tokens_received += 1
+
+    # ── Word extraction ──
+
+    def _pos_to_time(self, token_position: int) -> float:
+        """Convert token position to seconds."""
+        return token_position * self._seconds_per_token + self._global_time_offset
+
+    def _extract_new_words(self) -> List[ASRToken]:
+        """Extract complete words (all but the last, which may still be growing)."""
+        with self._text_lock:
+            text = self._accumulated_text
+        if not text:
+            return []
+
+        words = text.split()
+        new_words: List[ASRToken] = []
+        n_tokens = self._n_text_tokens_received
+        n_words_total = len(words)
+
+        while len(words) > self._n_committed_words + 1:
+            word = words[self._n_committed_words]
+            word_idx = self._n_committed_words
+
+            tok_start = int(word_idx / n_words_total * n_tokens) if n_words_total > 0 else 0
+            tok_end = int((word_idx + 1) / n_words_total * n_tokens) if n_words_total > 0 else 0
+
+            start_time = self._pos_to_time(tok_start)
+            end_time = self._pos_to_time(tok_end)
+
+            text_out = word if self._n_committed_words == 0 else " " + word
+            new_words.append(ASRToken(start=start_time, end=end_time, text=text_out))
+            self._n_committed_words += 1
+
+        return new_words
+
+    def _flush_all_pending_words(self) -> List[ASRToken]:
+        """Flush ALL words including the last partial one."""
+        with self._text_lock:
+            text = self._accumulated_text
+        if not text:
+            return []
+
+        words = text.split()
+        new_words: List[ASRToken] = []
+        n_tokens = max(self._n_text_tokens_received, 1)
+        n_words_total = max(len(words), 1)
+
+        while self._n_committed_words < len(words):
+            word = words[self._n_committed_words]
+            word_idx = self._n_committed_words
+
+            tok_start = int(word_idx / n_words_total * n_tokens)
+            tok_end = int((word_idx + 1) / n_words_total * n_tokens)
+
+            start_time = self._pos_to_time(tok_start)
+            end_time = self._pos_to_time(tok_end)
+
+            text_out = word if self._n_committed_words == 0 else " " + word
+            new_words.append(ASRToken(start=start_time, end=end_time, text=text_out))
+            self._n_committed_words += 1
+
+        return new_words
+
+    # ── Core processing ──
+
+    def _process_iter_inner(self, is_last: bool) -> Tuple[List[ASRToken], float]:
+        # Start generate thread when enough audio is buffered
+        if not self._generate_started:
+            if len(self._pending_audio) >= self._first_chunk_samples:
+                self._start_generate_thread()
+                self._feed_pending_audio()
+            else:
+                return [], self.end
+
+        # Feed any new pending audio
+        if self._generate_started and not self._generate_finished:
+            self._feed_pending_audio()
+
+        # If generate finished unexpectedly (EOS) but new audio arrived, restart
+        if self._generate_finished and len(self._pending_audio) >= self._first_chunk_samples:
+            self._drain_streamer()
+            flush_words = self._flush_all_pending_words()
+            # Reset for new utterance
+            old_offset = self._global_time_offset
+            self._reset_state()
+            self._global_time_offset = old_offset
+            self._start_generate_thread()
+            self._feed_pending_audio()
+            return flush_words, self.end
+
+        # Drain available text from streamer
+        self._drain_streamer()
+
+        # Extract complete words
+        new_words = self._extract_new_words()
+
+        if new_words:
+            logger.info(f"[voxtral-hf] returning {len(new_words)} words: {[w.text for w in new_words]}")
+
+        self.buffer = []
+        return new_words, self.end

whisperlivekit/voxtral_mlx/__init__.py

@@ -0,0 +1,6 @@
+"""Pure-MLX Voxtral Realtime backend for WhisperLiveKit."""
+
+from .loader import load_voxtral_model
+from .model import VoxtralMLXModel
+
+__all__ = ["load_voxtral_model", "VoxtralMLXModel"]

whisperlivekit/voxtral_mlx/loader.py

@@ -0,0 +1,282 @@
+"""
+Model weight loading for the MLX Voxtral Realtime backend.
+
+Supports two on-disk formats:
+  1. **Converted** (``config.json`` + ``model.safetensors``): ready-to-load,
+     with optional quantisation metadata.
+  2. **Original Mistral** (``params.json`` + ``consolidated.safetensors``):
+     requires weight renaming and conv-weight transposition.
+
+The public entry point is :func:`load_voxtral_model` which returns the
+model, tokenizer, and raw config dict.
+"""
+
+import json
+import logging
+import re
+from pathlib import Path
+
+import mlx.core as mx
+import mlx.nn as nn
+from huggingface_hub import snapshot_download
+
+from .model import VoxtralMLXModel
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_MODEL_ID = "mlx-community/Voxtral-Mini-4B-Realtime-6bit"
+
+# ---------------------------------------------------------------------------
+# Downloading
+# ---------------------------------------------------------------------------
+
+_ALLOWED_PATTERNS = [
+    "consolidated.safetensors",
+    "model*.safetensors",
+    "model.safetensors.index.json",
+    "params.json",
+    "config.json",
+    "tekken.json",
+]
+
+
+def download_weights(model_id: str = DEFAULT_MODEL_ID) -> Path:
+    """Download model files from HuggingFace Hub and return the local path."""
+    return Path(snapshot_download(model_id, allow_patterns=_ALLOWED_PATTERNS))
+
+
+# ---------------------------------------------------------------------------
+# Weight name remapping (Mistral → our naming)
+# ---------------------------------------------------------------------------
+
+_NAME_RULES: list[tuple[str, str]] = [
+    # Encoder convolutions
+    (r"whisper_encoder\.conv_layers\.0\.conv\.(.*)", r"encoder.conv1.\1"),
+    (r"whisper_encoder\.conv_layers\.1\.conv\.(.*)", r"encoder.conv2.\1"),
+    # Encoder transformer blocks
+    (r"whisper_encoder\.transformer\.layers\.(\d+)\.attention\.wq\.(.*)",
+     r"encoder.blocks.\1.self_attn.q_proj.\2"),
+    (r"whisper_encoder\.transformer\.layers\.(\d+)\.attention\.wk\.(.*)",
+     r"encoder.blocks.\1.self_attn.k_proj.\2"),
+    (r"whisper_encoder\.transformer\.layers\.(\d+)\.attention\.wv\.(.*)",
+     r"encoder.blocks.\1.self_attn.v_proj.\2"),
+    (r"whisper_encoder\.transformer\.layers\.(\d+)\.attention\.wo\.(.*)",
+     r"encoder.blocks.\1.self_attn.out_proj.\2"),
+    (r"whisper_encoder\.transformer\.layers\.(\d+)\.attention_norm\.(.*)",
+     r"encoder.blocks.\1.pre_attn_norm.\2"),
+    (r"whisper_encoder\.transformer\.layers\.(\d+)\.feed_forward\.w1\.(.*)",
+     r"encoder.blocks.\1.ffn.gate.\2"),
+    (r"whisper_encoder\.transformer\.layers\.(\d+)\.feed_forward\.w2\.(.*)",
+     r"encoder.blocks.\1.ffn.down.\2"),
+    (r"whisper_encoder\.transformer\.layers\.(\d+)\.feed_forward\.w3\.(.*)",
+     r"encoder.blocks.\1.ffn.up.\2"),
+    (r"whisper_encoder\.transformer\.layers\.(\d+)\.ffn_norm\.(.*)",
+     r"encoder.blocks.\1.pre_ffn_norm.\2"),
+    (r"whisper_encoder\.transformer\.norm\.(.*)", r"encoder.final_norm.\1"),
+    # Adapter
+    (r"audio_language_projection\.0\.weight", r"adapter.linear1.weight"),
+    (r"audio_language_projection\.2\.weight", r"adapter.linear2.weight"),
+    # Decoder embedding
+    (r"tok_embeddings\.weight", r"decoder.token_embedding.weight"),
+    # Decoder blocks
+    (r"layers\.(\d+)\.attention\.wq\.weight",
+     r"decoder.blocks.\1.self_attn.q_proj.weight"),
+    (r"layers\.(\d+)\.attention\.wk\.weight",
+     r"decoder.blocks.\1.self_attn.k_proj.weight"),
+    (r"layers\.(\d+)\.attention\.wv\.weight",
+     r"decoder.blocks.\1.self_attn.v_proj.weight"),
+    (r"layers\.(\d+)\.attention\.wo\.weight",
+     r"decoder.blocks.\1.self_attn.out_proj.weight"),
+    (r"layers\.(\d+)\.attention_norm\.weight",
+     r"decoder.blocks.\1.pre_attn_norm.weight"),
+    (r"layers\.(\d+)\.feed_forward\.w1\.weight",
+     r"decoder.blocks.\1.ffn.gate.weight"),
+    (r"layers\.(\d+)\.feed_forward\.w2\.weight",
+     r"decoder.blocks.\1.ffn.down.weight"),
+    (r"layers\.(\d+)\.feed_forward\.w3\.weight",
+     r"decoder.blocks.\1.ffn.up.weight"),
+    (r"layers\.(\d+)\.ffn_norm\.weight",
+     r"decoder.blocks.\1.pre_ffn_norm.weight"),
+    (r"layers\.(\d+)\.ada_rms_norm_t_cond\.0\.weight",
+     r"decoder.blocks.\1.adaptive_scale.proj_in.weight"),
+    (r"layers\.(\d+)\.ada_rms_norm_t_cond\.2\.weight",
+     r"decoder.blocks.\1.adaptive_scale.proj_out.weight"),
+    # Decoder final norm
+    (r"norm\.weight", r"decoder.final_norm.weight"),
+]
+
+_PREFIX_STRIP = re.compile(
+    r"^(mm_streams_embeddings\.embedding_module|mm_whisper_embeddings)\."
+)
+
+
+def _translate_weight_name(name: str) -> str | None:
+    name = _PREFIX_STRIP.sub("", name)
+    for pattern, replacement in _NAME_RULES:
+        result, n = re.subn(f"^{pattern}$", replacement, name)
+        if n:
+            return result
+    return None
+
+
+def _is_conv_weight(name: str) -> bool:
+    return ("conv1.weight" in name or "conv2.weight" in name) and "bias" not in name
+
+
+# ---------------------------------------------------------------------------
+# Converted-format weight remapping (voxmlx names → our names)
+# ---------------------------------------------------------------------------
+
+_CONVERTED_RULES: list[tuple[str, str]] = [
+    # Adapter
+    (r"adapter\.w_in\.(.*)", r"adapter.linear1.\1"),
+    (r"adapter\.w_out\.(.*)", r"adapter.linear2.\1"),
+    # Encoder transformer blocks
+    (r"encoder\.layers\.(\d+)\.attention\.(.*)", r"encoder.blocks.\1.self_attn.\2"),
+    (r"encoder\.layers\.(\d+)\.attn_norm\.(.*)", r"encoder.blocks.\1.pre_attn_norm.\2"),
+    (r"encoder\.layers\.(\d+)\.mlp\.gate_proj\.(.*)", r"encoder.blocks.\1.ffn.gate.\2"),
+    (r"encoder\.layers\.(\d+)\.mlp\.down_proj\.(.*)", r"encoder.blocks.\1.ffn.down.\2"),
+    (r"encoder\.layers\.(\d+)\.mlp\.up_proj\.(.*)", r"encoder.blocks.\1.ffn.up.\2"),
+    (r"encoder\.layers\.(\d+)\.ffn_norm\.(.*)", r"encoder.blocks.\1.pre_ffn_norm.\2"),
+    (r"encoder\.norm\.(.*)", r"encoder.final_norm.\1"),
+    # Decoder embedding
+    (r"language_model\.embed_tokens\.(.*)", r"decoder.token_embedding.\1"),
+    # Decoder blocks
+    (r"language_model\.layers\.(\d+)\.attention\.(.*)", r"decoder.blocks.\1.self_attn.\2"),
+    (r"language_model\.layers\.(\d+)\.attn_norm\.(.*)", r"decoder.blocks.\1.pre_attn_norm.\2"),
+    (r"language_model\.layers\.(\d+)\.mlp\.gate_proj\.(.*)", r"decoder.blocks.\1.ffn.gate.\2"),
+    (r"language_model\.layers\.(\d+)\.mlp\.down_proj\.(.*)", r"decoder.blocks.\1.ffn.down.\2"),
+    (r"language_model\.layers\.(\d+)\.mlp\.up_proj\.(.*)", r"decoder.blocks.\1.ffn.up.\2"),
+    (r"language_model\.layers\.(\d+)\.ffn_norm\.(.*)", r"decoder.blocks.\1.pre_ffn_norm.\2"),
+    (r"language_model\.layers\.(\d+)\.ada_norm\.linear_in\.(.*)",
+     r"decoder.blocks.\1.adaptive_scale.proj_in.\2"),
+    (r"language_model\.layers\.(\d+)\.ada_norm\.linear_out\.(.*)",
+     r"decoder.blocks.\1.adaptive_scale.proj_out.\2"),
+    (r"language_model\.norm\.(.*)", r"decoder.final_norm.\1"),
+]
+
+# Also remap o_proj → out_proj in both encoder and decoder
+_POST_RENAME = [
+    (r"\.o_proj\.", r".out_proj."),
+]
+
+
+def _remap_converted_name(name: str) -> str:
+    """Translate a converted-format weight name to our naming convention."""
+    for pattern, replacement in _CONVERTED_RULES:
+        result, n = re.subn(f"^{pattern}$", replacement, name)
+        if n:
+            name = result
+            break
+    for pattern, replacement in _POST_RENAME:
+        name = re.sub(pattern, replacement, name)
+    return name
+
+
+# ---------------------------------------------------------------------------
+# Loading strategies
+# ---------------------------------------------------------------------------
+
+def _has_converted_layout(path: Path) -> bool:
+    return (path / "config.json").exists() and not (path / "consolidated.safetensors").exists()
+
+
+def _load_converted_weights(path: Path):
+    with open(path / "config.json") as f:
+        config = json.load(f)
+
+    model = VoxtralMLXModel(config)
+
+    quant = config.get("quantization")
+    if quant is not None:
+        gs = quant["group_size"]
+        nn.quantize(
+            model,
+            group_size=gs,
+            bits=quant["bits"],
+            class_predicate=lambda _p, m: (
+                hasattr(m, "to_quantized") and m.weight.shape[-1] % gs == 0
+            ),
+        )
+
+    index_file = path / "model.safetensors.index.json"
+    if index_file.exists():
+        with open(index_file) as f:
+            shard_map = json.load(f)
+        shard_files = sorted(set(shard_map["weight_map"].values()))
+        weights = {}
+        for sf in shard_files:
+            weights.update(mx.load(str(path / sf)))
+    else:
+        weights = mx.load(str(path / "model.safetensors"))
+
+    remapped = {_remap_converted_name(k): v for k, v in weights.items()}
+    model.load_weights(list(remapped.items()))
+    mx.eval(model.parameters())
+    return model, config
+
+
+def _load_original_weights(path: Path):
+    with open(path / "params.json") as f:
+        config = json.load(f)
+
+    model = VoxtralMLXModel(config)
+
+    raw = mx.load(str(path / "consolidated.safetensors"))
+    mapped: dict[str, mx.array] = {}
+    skipped: list[str] = []
+
+    for name, tensor in raw.items():
+        if name == "output.weight":
+            continue
+        new_name = _translate_weight_name(name)
+        if new_name is None:
+            skipped.append(name)
+            continue
+        # Conv weights: PyTorch [C_out, C_in, K] → MLX [C_out, K, C_in]
+        if _is_conv_weight(new_name):
+            tensor = mx.swapaxes(tensor, 1, 2)
+        mapped[new_name] = tensor
+
+    if skipped:
+        logger.warning("Skipped %d unrecognised weight keys (first 5: %s)", len(skipped), skipped[:5])
+
+    model.load_weights(list(mapped.items()))
+    mx.eval(model.parameters())
+    return model, config
+
+
+# ---------------------------------------------------------------------------
+# Tokenizer
+# ---------------------------------------------------------------------------
+
+def _load_tokenizer(model_dir: Path):
+    from mistral_common.tokens.tokenizers.tekken import Tekkenizer
+    return Tekkenizer.from_file(str(model_dir / "tekken.json"))
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+def load_voxtral_model(path_or_id: str = DEFAULT_MODEL_ID):
+    """Load a Voxtral Realtime model and its tokenizer.
+
+    Args:
+        path_or_id: Local directory path **or** a HuggingFace model ID.
+
+    Returns:
+        ``(model, tokenizer, config)``
+    """
+    p = Path(path_or_id)
+    if not p.exists():
+        p = download_weights(path_or_id)
+
+    if _has_converted_layout(p):
+        model, config = _load_converted_weights(p)
+    else:
+        model, config = _load_original_weights(p)
+
+    tokenizer = _load_tokenizer(p)
+    logger.info("Voxtral MLX model loaded from %s", p)
+    return model, tokenizer, config

whisperlivekit/voxtral_mlx/model.py

@@ -0,0 +1,534 @@
+"""
+Voxtral Realtime MLX model — encoder, decoder, adapter, and top-level model.
+
+Architecture:
+    audio → StreamingEncoder → EncoderToDecoderAdapter → TextDecoder → logits
+    with DelayEmbedding providing time-conditioning to the decoder.
+
+The model supports both batch inference (full audio) and incremental streaming
+(one chunk at a time with cached encoder/decoder state).
+"""
+
+import math
+
+import mlx.core as mx
+import mlx.nn as nn
+
+
+# ---------------------------------------------------------------------------
+# KV Cache
+# ---------------------------------------------------------------------------
+
+
+class SlidingKVCache:
+    """Bounded key-value cache with rotating buffer for sliding-window attention.
+
+    Uses in-place writes for single-token autoregressive steps and
+    concatenation for multi-token prefills. Pre-allocates in blocks of
+    ``alloc_step`` entries to reduce repeated allocation.
+    """
+
+    alloc_step = 256
+
+    def __init__(self, capacity: int):
+        self.capacity = capacity
+        self.keys = None
+        self.values = None
+        self._offset = 0
+        self._write_idx = 0
+
+    @property
+    def offset(self) -> int:
+        return self._offset
+
+    # -- helpers --
+
+    def _reorder(self, buf):
+        """Return *buf* in temporal order (unwrap the circular buffer)."""
+        if self._write_idx == buf.shape[2]:
+            return buf
+        if self._write_idx < self._offset:
+            return mx.concatenate(
+                [buf[..., self._write_idx:, :], buf[..., : self._write_idx, :]],
+                axis=2,
+            )
+        return buf[..., : self._write_idx, :]
+
+    def _drop_oldest(self, buf, n_drop, tail=None):
+        parts = [buf[..., n_drop:, :]] if n_drop > 0 else [buf]
+        if tail is not None:
+            parts.append(tail)
+        return mx.concatenate(parts, axis=2)
+
+    # -- update strategies --
+
+    def _append_concat(self, k, v):
+        """Multi-token update via concatenation (used during prefill)."""
+        if self.keys is None:
+            self.keys, self.values = k, v
+        else:
+            self.keys = self._reorder(self.keys)
+            self.values = self._reorder(self.values)
+            self._write_idx = self.keys.shape[2]
+            overflow = self._write_idx - self.capacity + 1
+            self.keys = self._drop_oldest(self.keys, overflow, k)
+            self.values = self._drop_oldest(self.values, overflow, v)
+        self._offset += k.shape[2]
+        self._write_idx = self.keys.shape[2]
+        return self.keys, self.values
+
+    def _write_inplace(self, k, v):
+        """Single-token update via in-place write (autoregressive step)."""
+        B, n_heads, S, dim_k = k.shape
+        dim_v = v.shape[3]
+        prev = self._offset
+
+        if self.keys is None or (
+            prev >= self.keys.shape[2] and self.keys.shape[2] < self.capacity
+        ):
+            n_new = min(self.alloc_step, self.capacity - prev)
+            fresh_k = mx.zeros((B, n_heads, n_new, dim_k), k.dtype)
+            fresh_v = mx.zeros((B, n_heads, n_new, dim_v), v.dtype)
+            if self.keys is not None:
+                self.keys = mx.concatenate([self.keys, fresh_k], axis=2)
+                self.values = mx.concatenate([self.values, fresh_v], axis=2)
+            else:
+                self.keys, self.values = fresh_k, fresh_v
+            self._write_idx = prev
+
+        overflow = self.keys.shape[2] - self.capacity
+        if overflow > 0:
+            self.keys = self._drop_oldest(self.keys, overflow)
+            self.values = self._drop_oldest(self.values, overflow)
+            self._write_idx = self.capacity
+
+        if self._write_idx == self.capacity:
+            self._write_idx = 0
+
+        self.keys[..., self._write_idx : self._write_idx + S, :] = k
+        self.values[..., self._write_idx : self._write_idx + S, :] = v
+        self._offset += S
+        self._write_idx += S
+
+        if self._offset < self.capacity:
+            return (
+                self.keys[..., : self._offset, :],
+                self.values[..., : self._offset, :],
+            )
+        return self.keys, self.values
+
+    # -- public API --
+
+    def update_and_fetch(self, k, v):
+        if k.shape[2] == 1:
+            return self._write_inplace(k, v)
+        return self._append_concat(k, v)
+
+
+# ---------------------------------------------------------------------------
+# Encoder components
+# ---------------------------------------------------------------------------
+
+
+class CausalConv(nn.Module):
+    """1-D causal convolution (left-padded so no future leakage)."""
+
+    def __init__(self, channels_in: int, channels_out: int, kernel: int, stride: int = 1):
+        super().__init__()
+        self.stride = stride
+        self.kernel = kernel
+        self.left_pad = kernel - stride
+        self.weight = mx.zeros((channels_out, kernel, channels_in))
+        self.bias = mx.zeros((channels_out,))
+
+    def __call__(self, x: mx.array) -> mx.array:
+        if self.left_pad > 0:
+            x = mx.pad(x, [(0, 0), (self.left_pad, 0), (0, 0)])
+        return mx.conv1d(x, self.weight, stride=self.stride) + self.bias
+
+
+class _EncoderSelfAttention(nn.Module):
+    def __init__(self, dim: int, n_heads: int, head_dim: int, rope_theta: float):
+        super().__init__()
+        self.n_heads = n_heads
+        self.head_dim = head_dim
+        self.scale = head_dim**-0.5
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=True)
+        self.k_proj = nn.Linear(dim, n_heads * head_dim, bias=False)
+        self.v_proj = nn.Linear(dim, n_heads * head_dim, bias=True)
+        self.out_proj = nn.Linear(n_heads * head_dim, dim, bias=True)
+        self.rope_theta = rope_theta
+
+    def __call__(self, x, mask, cache=None):
+        B, L, _ = x.shape
+        q = self.q_proj(x).reshape(B, L, self.n_heads, self.head_dim).transpose(0, 2, 1, 3)
+        k = self.k_proj(x).reshape(B, L, self.n_heads, self.head_dim).transpose(0, 2, 1, 3)
+        v = self.v_proj(x).reshape(B, L, self.n_heads, self.head_dim).transpose(0, 2, 1, 3)
+
+        pos = cache.offset if cache is not None else 0
+        q = mx.fast.rope(q, self.head_dim, traditional=True, base=self.rope_theta, scale=1.0, offset=pos)
+        k = mx.fast.rope(k, self.head_dim, traditional=True, base=self.rope_theta, scale=1.0, offset=pos)
+
+        if cache is not None:
+            k, v = cache.update_and_fetch(k, v)
+
+        out = mx.fast.scaled_dot_product_attention(q, k, v, scale=self.scale, mask=mask)
+        return self.out_proj(out.transpose(0, 2, 1, 3).reshape(B, L, -1))
+
+
+class _EncoderFFN(nn.Module):
+    """SwiGLU feed-forward for encoder layers."""
+
+    def __init__(self, dim: int, hidden: int):
+        super().__init__()
+        self.gate = nn.Linear(dim, hidden, bias=False)
+        self.up = nn.Linear(dim, hidden, bias=False)
+        self.down = nn.Linear(hidden, dim, bias=True)
+
+    def __call__(self, x):
+        return self.down(nn.silu(self.gate(x)) * self.up(x))
+
+
+class _EncoderBlock(nn.Module):
+    def __init__(self, dim, n_heads, head_dim, hidden, rope_theta):
+        super().__init__()
+        self.pre_attn_norm = nn.RMSNorm(dim, eps=1e-5)
+        self.self_attn = _EncoderSelfAttention(dim, n_heads, head_dim, rope_theta)
+        self.pre_ffn_norm = nn.RMSNorm(dim, eps=1e-5)
+        self.ffn = _EncoderFFN(dim, hidden)
+
+    def __call__(self, x, mask, cache=None):
+        x = x + self.self_attn(self.pre_attn_norm(x), mask, cache=cache)
+        x = x + self.ffn(self.pre_ffn_norm(x))
+        return x
+
+
+class StreamingEncoder(nn.Module):
+    """Causal Whisper-style encoder with two causal convolutions followed by
+    a stack of transformer blocks.  Supports both full-sequence and
+    incremental (streaming) forward passes."""
+
+    def __init__(
+        self,
+        mel_channels: int = 128,
+        dim: int = 1280,
+        n_layers: int = 32,
+        n_heads: int = 32,
+        head_dim: int = 64,
+        hidden_dim: int = 5120,
+        rope_theta: float = 1e6,
+        sliding_window: int = 750,
+    ):
+        super().__init__()
+        self.conv1 = CausalConv(mel_channels, dim, kernel=3, stride=1)
+        self.conv2 = CausalConv(dim, dim, kernel=3, stride=2)
+        self.blocks = [
+            _EncoderBlock(dim, n_heads, head_dim, hidden_dim, rope_theta)
+            for _ in range(n_layers)
+        ]
+        self.final_norm = nn.RMSNorm(dim, eps=1e-5)
+        self.sliding_window = sliding_window
+
+    # -- full-sequence --
+
+    def _apply_convs(self, mel: mx.array) -> mx.array:
+        x = mel.T[None, :, :]  # [1, T, mel_channels]
+        x = nn.gelu(self.conv1(x))
+        x = nn.gelu(self.conv2(x))
+        return x
+
+    def forward(self, mel: mx.array) -> mx.array:
+        x = self._apply_convs(mel.astype(self.conv1.weight.dtype))
+        for blk in self.blocks:
+            x = blk(x, mask="causal")
+        return self.final_norm(x)
+
+    # -- incremental (streaming) --
+
+    def forward_conv_incremental(self, x_in, tail1, tail2):
+        """Process new mel frames through the two causal convs using cached tails.
+
+        Args:
+            x_in: [1, N, mel_channels]
+            tail1: [1, pad1, mel_channels] or None (first call)
+            tail2: [1, pad2, dim] or None (first call)
+
+        Returns:
+            (out, new_tail1, new_tail2)
+        """
+        # Conv1 (kernel=3, stride=1 → left_pad=2)
+        if tail1 is not None:
+            c1_in = mx.concatenate([tail1, x_in], axis=1)
+        else:
+            c1_in = mx.pad(x_in, [(0, 0), (self.conv1.left_pad, 0), (0, 0)])
+        new_tail1 = x_in[:, -self.conv1.left_pad :, :]
+        c1_out = nn.gelu(
+            mx.conv1d(c1_in, self.conv1.weight, stride=self.conv1.stride) + self.conv1.bias
+        )
+
+        # Conv2 (kernel=3, stride=2 → left_pad=1)
+        if tail2 is not None:
+            c2_in = mx.concatenate([tail2, c1_out], axis=1)
+        else:
+            c2_in = mx.pad(c1_out, [(0, 0), (self.conv2.left_pad, 0), (0, 0)])
+        new_tail2 = c1_out[:, -self.conv2.left_pad :, :]
+        c2_out = nn.gelu(
+            mx.conv1d(c2_in, self.conv2.weight, stride=self.conv2.stride) + self.conv2.bias
+        )
+
+        return c2_out, new_tail1, new_tail2
+
+    def forward_transformer_incremental(self, x, cache_list):
+        """Run transformer blocks with per-layer KV caches."""
+        for i, blk in enumerate(self.blocks):
+            x = blk(x, mask="causal", cache=cache_list[i])
+        return self.final_norm(x)
+
+
+# ---------------------------------------------------------------------------
+# Decoder components
+# ---------------------------------------------------------------------------
+
+
+class _DecoderAttention(nn.Module):
+    """Grouped-query attention for the text decoder."""
+
+    def __init__(self, dim, n_heads, n_kv_heads, head_dim, rope_theta):
+        super().__init__()
+        self.n_heads = n_heads
+        self.n_kv_heads = n_kv_heads
+        self.head_dim = head_dim
+        self.scale = head_dim**-0.5
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=False)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.out_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+        self.rope_theta = rope_theta
+
+    def __call__(self, x, mask=None, cache=None):
+        B, L, _ = x.shape
+        q = self.q_proj(x).reshape(B, L, self.n_heads, self.head_dim).transpose(0, 2, 1, 3)
+        k = self.k_proj(x).reshape(B, L, self.n_kv_heads, self.head_dim).transpose(0, 2, 1, 3)
+        v = self.v_proj(x).reshape(B, L, self.n_kv_heads, self.head_dim).transpose(0, 2, 1, 3)
+
+        pos = cache.offset if cache is not None else 0
+        q = mx.fast.rope(q, self.head_dim, traditional=True, base=self.rope_theta, scale=1.0, offset=pos)
+        k = mx.fast.rope(k, self.head_dim, traditional=True, base=self.rope_theta, scale=1.0, offset=pos)
+
+        if cache is not None:
+            k, v = cache.update_and_fetch(k, v)
+
+        out = mx.fast.scaled_dot_product_attention(q, k, v, scale=self.scale, mask=mask)
+        return self.out_proj(out.transpose(0, 2, 1, 3).reshape(B, L, -1))
+
+
+class _DecoderFFN(nn.Module):
+    """SwiGLU feed-forward for decoder layers."""
+
+    def __init__(self, dim, hidden):
+        super().__init__()
+        self.gate = nn.Linear(dim, hidden, bias=False)
+        self.up = nn.Linear(dim, hidden, bias=False)
+        self.down = nn.Linear(hidden, dim, bias=False)
+
+    def __call__(self, x):
+        return self.down(nn.silu(self.gate(x)) * self.up(x))
+
+
+class AdaptiveScaling(nn.Module):
+    """Small MLP that produces a multiplicative scale from the delay embedding,
+    used to condition the FFN on the streaming delay."""
+
+    def __init__(self, dim, bottleneck):
+        super().__init__()
+        self.proj_in = nn.Linear(dim, bottleneck, bias=False)
+        self.proj_out = nn.Linear(bottleneck, dim, bias=False)
+
+    def __call__(self, cond):
+        return self.proj_out(nn.gelu(self.proj_in(cond)))
+
+
+class _DecoderBlock(nn.Module):
+    def __init__(self, dim, n_heads, n_kv_heads, head_dim, hidden, rope_theta, cond_dim):
+        super().__init__()
+        self.pre_attn_norm = nn.RMSNorm(dim, eps=1e-5)
+        self.self_attn = _DecoderAttention(dim, n_heads, n_kv_heads, head_dim, rope_theta)
+        self.adaptive_scale = AdaptiveScaling(dim, cond_dim)
+        self.pre_ffn_norm = nn.RMSNorm(dim, eps=1e-5)
+        self.ffn = _DecoderFFN(dim, hidden)
+
+    def __call__(self, x, delay_cond, mask=None, cache=None):
+        x = x + self.self_attn(self.pre_attn_norm(x), mask, cache)
+        scaled = self.pre_ffn_norm(x) * (1.0 + self.adaptive_scale(delay_cond))
+        x = x + self.ffn(scaled)
+        return x
+
+
+class TextDecoder(nn.Module):
+    """Mistral-style causal language model with adaptive time-conditioning."""
+
+    def __init__(
+        self,
+        dim: int = 3072,
+        n_layers: int = 26,
+        n_heads: int = 32,
+        n_kv_heads: int = 8,
+        head_dim: int = 128,
+        hidden_dim: int = 9216,
+        vocab_size: int = 131072,
+        rope_theta: float = 1e6,
+        cond_dim: int = 32,
+    ):
+        super().__init__()
+        self.token_embedding = nn.Embedding(vocab_size, dim)
+        self.blocks = [
+            _DecoderBlock(dim, n_heads, n_kv_heads, head_dim, hidden_dim, rope_theta, cond_dim)
+            for _ in range(n_layers)
+        ]
+        self.final_norm = nn.RMSNorm(dim, eps=1e-5)
+
+    def embed(self, token_ids: mx.array) -> mx.array:
+        return self.token_embedding(token_ids)
+
+    def __call__(self, x, delay_cond, mask=None, cache=None):
+        delay_cond = delay_cond.astype(x.dtype)
+        for i, blk in enumerate(self.blocks):
+            blk_cache = cache[i] if cache is not None else None
+            x = blk(x, delay_cond, mask, blk_cache)
+        x = self.final_norm(x)
+        return self.token_embedding.as_linear(x)
+
+
+# ---------------------------------------------------------------------------
+# Adapter & embeddings
+# ---------------------------------------------------------------------------
+
+
+class EncoderToDecoderAdapter(nn.Module):
+    """Two-layer projection from encoder space to decoder space."""
+
+    def __init__(self, enc_dim: int, dec_dim: int):
+        super().__init__()
+        self.linear1 = nn.Linear(enc_dim, dec_dim, bias=False)
+        self.linear2 = nn.Linear(dec_dim, dec_dim, bias=False)
+
+    def __call__(self, x):
+        return self.linear2(nn.gelu(self.linear1(x)))
+
+
+class DelayEmbedding(nn.Module):
+    """Sinusoidal embedding that encodes the streaming delay as a conditioning
+    vector for the decoder's adaptive scaling."""
+
+    def __init__(self, dim: int = 3072, theta: float = 10000.0):
+        super().__init__()
+        self.dim = dim
+        half = dim // 2
+        freqs = mx.exp(-math.log(theta) * mx.arange(half, dtype=mx.float32) / half)
+        self._freqs = freqs
+
+    def __call__(self, delay: mx.array) -> mx.array:
+        t = delay.reshape(-1, 1).astype(mx.float32)
+        angles = t * self._freqs
+        return mx.concatenate([mx.cos(angles), mx.sin(angles)], axis=-1)
+
+
+# ---------------------------------------------------------------------------
+# Top-level model
+# ---------------------------------------------------------------------------
+
+
+class VoxtralMLXModel(nn.Module):
+    """Top-level Voxtral Realtime model wiring encoder, adapter, and decoder."""
+
+    def __init__(self, config: dict):
+        super().__init__()
+
+        enc_cfg = config["multimodal"]["whisper_model_args"]["encoder_args"]
+        audio_cfg = enc_cfg["audio_encoding_args"]
+        ds_factor = config["multimodal"]["whisper_model_args"]["downsample_args"]["downsample_factor"]
+
+        self.encoder = StreamingEncoder(
+            mel_channels=audio_cfg["num_mel_bins"],
+            dim=enc_cfg["dim"],
+            n_layers=enc_cfg["n_layers"],
+            n_heads=enc_cfg["n_heads"],
+            head_dim=enc_cfg["head_dim"],
+            hidden_dim=enc_cfg["hidden_dim"],
+            rope_theta=enc_cfg["rope_theta"],
+            sliding_window=enc_cfg["sliding_window"],
+        )
+
+        adapter_input_dim = enc_cfg["dim"] * ds_factor
+        decoder_dim = config["dim"]
+        cond_bottleneck = config.get("ada_rms_norm_t_cond_dim", 32)
+
+        self.adapter = EncoderToDecoderAdapter(adapter_input_dim, decoder_dim)
+
+        self.decoder = TextDecoder(
+            dim=decoder_dim,
+            n_layers=config["n_layers"],
+            n_heads=config["n_heads"],
+            n_kv_heads=config["n_kv_heads"],
+            head_dim=config["head_dim"],
+            hidden_dim=config["hidden_dim"],
+            vocab_size=config["vocab_size"],
+            rope_theta=config["rope_theta"],
+            cond_dim=cond_bottleneck,
+        )
+
+        self.delay_embedding = DelayEmbedding(dim=decoder_dim)
+        self.ds_factor = ds_factor
+
+    # -- batch encode --
+
+    def encode(self, mel: mx.array) -> mx.array:
+        T = mel.shape[1]
+        if T % 2 != 0:
+            mel = mel[:, 1:]
+
+        h = self.encoder.forward(mel)  # [1, T/2, enc_dim]
+        h = h[0]
+
+        n = h.shape[0]
+        trim = n % self.ds_factor
+        if trim:
+            h = h[trim:]
+            n = h.shape[0]
+
+        h = h.reshape(n // self.ds_factor, -1)
+        return self.adapter(h)
+
+    # -- incremental encode --
+
+    def encode_incremental(self, new_mel, conv_tail1, conv_tail2, enc_cache, ds_remainder):
+        """Incrementally encode new mel frames.
+
+        Returns:
+            (audio_embeds | None, conv_tail1, conv_tail2, enc_cache, ds_remainder)
+        """
+        x = new_mel.T[None, :, :].astype(self.encoder.conv1.weight.dtype)
+
+        x, conv_tail1, conv_tail2 = self.encoder.forward_conv_incremental(x, conv_tail1, conv_tail2)
+
+        if enc_cache is None:
+            enc_cache = [SlidingKVCache(100_000) for _ in range(len(self.encoder.blocks))]
+
+        x = self.encoder.forward_transformer_incremental(x, enc_cache)
+        x = x[0]  # [N, enc_dim]
+
+        if ds_remainder is not None:
+            x = mx.concatenate([ds_remainder, x])
+
+        n_full = (x.shape[0] // self.ds_factor) * self.ds_factor
+        if n_full == 0:
+            return None, conv_tail1, conv_tail2, enc_cache, x
+
+        leftover = x[n_full:] if x.shape[0] > n_full else None
+        x = x[:n_full].reshape(n_full // self.ds_factor, -1)
+        return self.adapter(x), conv_tail1, conv_tail2, enc_cache, leftover
+
+    # -- decode --
+
+    def decode(self, embeddings, delay_cond, mask=None, cache=None):
+        return self.decoder(embeddings, delay_cond, mask, cache)

whisperlivekit/voxtral_mlx/spectrogram.py

@@ -0,0 +1,202 @@
+"""
+Mel spectrogram computation for Voxtral Realtime.
+
+Provides both a full-audio function and an incremental streaming variant
+that maintains overlap state between calls.  The DFT is computed via
+matrix multiplication in MLX — no external FFT dependency required.
+"""
+
+import math
+
+import mlx.core as mx
+import numpy as np
+
+# Audio / mel constants matching the Voxtral Realtime model expectations.
+SAMPLE_RATE = 16_000
+WINDOW_SIZE = 400        # n_fft
+HOP = 160
+MEL_BANDS = 128
+MEL_MAX = 1.5            # global log-mel normalisation ceiling
+# Each output audio token spans: hop * conv_stride(2) * downsample_factor(4)
+SAMPLES_PER_TOKEN = HOP * 2 * 4  # = 1280 samples = 80 ms
+
+# Padding tokens used by the model prompt structure.
+LEFT_PAD_TOKENS = 32
+RIGHT_PAD_TOKENS = 17
+
+
+# ---------------------------------------------------------------------------
+# Slaney mel filterbank
+# ---------------------------------------------------------------------------
+
+def _build_slaney_filterbank(
+    sr: int = SAMPLE_RATE,
+    n_fft: int = WINDOW_SIZE,
+    n_mels: int = MEL_BANDS,
+    lo_hz: float = 0.0,
+    hi_hz: float = 8000.0,
+) -> np.ndarray:
+    """Compute a Slaney-normalised triangular mel filterbank.
+
+    Returns an array of shape ``[n_mels, n_fft//2 + 1]``.
+    """
+
+    def _hz2mel(f):
+        threshold = 1000.0
+        base_mel = 15.0
+        log_coeff = 27.0 / np.log(6.4)
+        mel = 3.0 * f / 200.0
+        if isinstance(f, np.ndarray):
+            above = f >= threshold
+            mel[above] = base_mel + np.log(f[above] / threshold) * log_coeff
+        elif f >= threshold:
+            mel = base_mel + np.log(f / threshold) * log_coeff
+        return mel
+
+    def _mel2hz(m):
+        threshold = 1000.0
+        base_mel = 15.0
+        log_coeff = np.log(6.4) / 27.0
+        hz = 200.0 * m / 3.0
+        above = m >= base_mel
+        hz[above] = threshold * np.exp(log_coeff * (m[above] - base_mel))
+        return hz
+
+    n_bins = n_fft // 2 + 1
+    fft_hz = np.linspace(0, sr / 2, n_bins)
+    mel_lo, mel_hi = _hz2mel(lo_hz), _hz2mel(hi_hz)
+    mel_pts = np.linspace(mel_lo, mel_hi, n_mels + 2)
+    hz_pts = _mel2hz(mel_pts)
+    diffs = np.diff(hz_pts)
+
+    slopes = np.expand_dims(hz_pts, 0) - np.expand_dims(fft_hz, 1)
+    rising = -slopes[:, :-2] / diffs[:-1]
+    falling = slopes[:, 2:] / diffs[1:]
+    fb = np.maximum(0.0, np.minimum(rising, falling))
+
+    # Slaney area normalisation
+    widths = 2.0 / (hz_pts[2 : n_mels + 2] - hz_pts[:n_mels])
+    fb *= np.expand_dims(widths, 0)
+    return fb.T.astype(np.float32)
+
+
+_CACHED_FILTERS: mx.array | None = None
+
+
+def _mel_filters() -> mx.array:
+    global _CACHED_FILTERS
+    if _CACHED_FILTERS is None:
+        _CACHED_FILTERS = mx.array(_build_slaney_filterbank())
+    return _CACHED_FILTERS
+
+
+# ---------------------------------------------------------------------------
+# DFT helpers
+# ---------------------------------------------------------------------------
+
+def _hann_window() -> mx.array:
+    return mx.array(np.hanning(WINDOW_SIZE + 1)[:-1].astype(np.float32))
+
+
+def _dft_matrices():
+    """Pre-compute the real / imaginary DFT basis matrices."""
+    n_bins = WINDOW_SIZE // 2 + 1
+    k = mx.arange(n_bins, dtype=mx.float32)[:, None]
+    n = mx.arange(WINDOW_SIZE, dtype=mx.float32)[None, :]
+    phase = -2.0 * math.pi * (k @ n) / WINDOW_SIZE
+    return mx.cos(phase), mx.sin(phase)
+
+
+def _stft_frames(audio: mx.array, window: mx.array) -> mx.array:
+    """Frame *audio* using the Hann window and compute power spectrogram."""
+    n_bins = WINDOW_SIZE // 2 + 1
+    n_frames = 1 + (audio.shape[0] - WINDOW_SIZE) // HOP
+    if n_frames <= 0:
+        return mx.zeros((0, n_bins))
+
+    offsets = (mx.arange(n_frames) * HOP)[:, None]
+    indices = offsets + mx.arange(WINDOW_SIZE)[None, :]
+    windowed = audio[indices] * window[None, :]
+
+    dft_re, dft_im = _dft_matrices()
+    real_part = windowed @ dft_re.T
+    imag_part = windowed @ dft_im.T
+    return real_part ** 2 + imag_part ** 2
+
+
+def _apply_mel_and_log(power: mx.array) -> mx.array:
+    """Convert a power spectrogram to log-mel and normalise."""
+    mel = power @ _mel_filters().T
+    log_mel = mx.log10(mx.maximum(mel, 1e-10))
+    log_mel = mx.maximum(log_mel, MEL_MAX - 8.0)
+    return (log_mel + 4.0) / 4.0
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+def compute_mel(audio: np.ndarray) -> mx.array:
+    """Compute log-mel spectrogram for a complete audio signal.
+
+    Args:
+        audio: 1-D float32 numpy array at ``SAMPLE_RATE``.
+
+    Returns:
+        ``[MEL_BANDS, T]`` MLX array.
+    """
+    x = mx.array(audio)
+    pad = WINDOW_SIZE // 2
+    x = mx.pad(x, [(pad, pad)])
+    window = _hann_window()
+
+    power = _stft_frames(x, window)
+    # Drop last frame to match reference STFT behaviour
+    power = power[:-1]
+    return _apply_mel_and_log(power).T
+
+
+def compute_mel_streaming(
+    chunk: np.ndarray,
+    overlap: np.ndarray | None,
+) -> tuple[mx.array, np.ndarray]:
+    """Incrementally compute log-mel for a new audio chunk.
+
+    Args:
+        chunk: New audio samples (float32 numpy).
+        overlap: The last ``WINDOW_SIZE - HOP`` = 240 samples from the
+            previous call, or *None* on the first call (uses zero-padding).
+
+    Returns:
+        ``(mel, new_overlap)`` where *mel* is ``[MEL_BANDS, N]`` and
+        *new_overlap* is the 240-sample tail for the next call.
+    """
+    tail_len = WINDOW_SIZE - HOP  # 240
+
+    if overlap is not None:
+        combined = np.concatenate([overlap, chunk])
+    else:
+        combined = np.concatenate([np.zeros(WINDOW_SIZE // 2, dtype=np.float32), chunk])
+
+    new_overlap = combined[-tail_len:].copy()
+
+    x = mx.array(combined)
+    window = _hann_window()
+    power = _stft_frames(x, window)
+
+    if power.shape[0] == 0:
+        return mx.zeros((MEL_BANDS, 0)), new_overlap
+
+    return _apply_mel_and_log(power).T, new_overlap
+
+
+def pad_audio(
+    audio: np.ndarray,
+    n_left: int = LEFT_PAD_TOKENS,
+    n_right: int = RIGHT_PAD_TOKENS,
+) -> np.ndarray:
+    """Pad audio with silence for batch (non-streaming) inference."""
+    left = n_left * SAMPLES_PER_TOKEN
+    align = (SAMPLES_PER_TOKEN - (len(audio) % SAMPLES_PER_TOKEN)) % SAMPLES_PER_TOKEN
+    right = align + n_right * SAMPLES_PER_TOKEN
+    return np.pad(audio, (left, right))

whisperlivekit/voxtral_mlx_asr.py

@@ -0,0 +1,521 @@
+"""
+Pure-MLX Voxtral Realtime ASR backend for WhisperLiveKit.
+
+Provides ``VoxtralMLXASR`` (model holder) and ``VoxtralMLXOnlineProcessor``
+(streaming processor) that plug into WhisperLiveKit's audio processing
+pipeline via ``insert_audio_chunk`` / ``process_iter`` / ``get_buffer`` etc.
+
+Unlike the HuggingFace backend, this runs the full inference loop in-process
+(no background thread / queue) — MLX operations on Apple Silicon are fast
+enough to run synchronously inside ``asyncio.to_thread(process_iter)``.
+"""
+
+import logging
+import sys
+import time
+from typing import List, Optional, Tuple
+
+import mlx.core as mx
+import numpy as np
+from mistral_common.tokens.tokenizers.base import SpecialTokenPolicy
+
+from whisperlivekit.timed_objects import ASRToken, Transcript
+from whisperlivekit.voxtral_mlx.loader import load_voxtral_model, DEFAULT_MODEL_ID
+from whisperlivekit.voxtral_mlx.model import SlidingKVCache
+from whisperlivekit.voxtral_mlx.spectrogram import (
+    SAMPLES_PER_TOKEN,
+    LEFT_PAD_TOKENS,
+    RIGHT_PAD_TOKENS,
+    compute_mel_streaming,
+)
+
+logger = logging.getLogger(__name__)
+
+# Decoder sliding-window size (matches the model's training configuration).
+_DECODER_WINDOW = 8192
+
+
+def _prompt_tokens(tokenizer, n_left_pad=LEFT_PAD_TOKENS, n_delay=6):
+    """Build the prompt token sequence and return ``(token_ids, n_delay)``."""
+    pad_id = tokenizer.get_special_token("[STREAMING_PAD]")
+    ids = [tokenizer.bos_id] + [pad_id] * (n_left_pad + n_delay)
+    return ids, n_delay
+
+
+# ---------------------------------------------------------------------------
+# Model holder
+# ---------------------------------------------------------------------------
+
+
+class VoxtralMLXASR:
+    """Lightweight model holder — loads the MLX Voxtral model once and keeps
+    it alive for the lifetime of the server."""
+
+    sep = " "
+    SAMPLING_RATE = 16_000
+
+    def __init__(self, logfile=sys.stderr, **kwargs):
+        self.logfile = logfile
+        self.transcribe_kargs = {}
+
+        lan = kwargs.get("lan", "auto")
+        self.original_language = None if lan == "auto" else lan
+
+        model_path = kwargs.get("model_dir") or kwargs.get("model_path")
+        if not model_path:
+            model_size = kwargs.get("model_size", "")
+            if model_size and ("/" in model_size or model_size.startswith(".")):
+                model_path = model_size
+            else:
+                model_path = DEFAULT_MODEL_ID
+
+        t0 = time.time()
+        logger.info("Loading Voxtral MLX model '%s' ...", model_path)
+        self.model, self.tokenizer, self.config = load_voxtral_model(model_path)
+        logger.info("Voxtral MLX model loaded in %.2fs", time.time() - t0)
+
+        self.backend_choice = "voxtral-mlx"
+
+    def transcribe(self, audio):
+        pass  # all work happens in the online processor
+
+
+# ---------------------------------------------------------------------------
+# Online processor
+# ---------------------------------------------------------------------------
+
+
+class VoxtralMLXOnlineProcessor:
+    """Streaming processor that incrementally encodes audio and decodes text
+    using the MLX Voxtral model.
+
+    Lifecycle (called by ``AudioProcessor.transcription_processor``):
+
+        insert_audio_chunk(pcm, time)  →  process_iter()  →  get_buffer()
+                      ... repeat ...
+        start_silence() / end_silence()
+        finish()
+    """
+
+    SAMPLING_RATE = 16_000
+
+    def __init__(self, asr: VoxtralMLXASR, logfile=sys.stderr):
+        self.asr = asr
+        self.logfile = logfile
+        self.end = 0.0
+        self.buffer: list = []
+        self.audio_buffer = np.array([], dtype=np.float32)
+
+        self._model = asr.model
+        self._tokenizer = asr.tokenizer
+
+        # Pre-compute prompt tokens and delay conditioning (constant across utterances).
+        self._prompt_ids, self._n_delay = _prompt_tokens(self._tokenizer)
+        self._prefix_len = len(self._prompt_ids)
+
+        self._delay_cond = self._model.delay_embedding(
+            mx.array([self._n_delay], dtype=mx.float32)
+        )
+        mx.eval(self._delay_cond)
+
+        self._prompt_embeds = self._model.decoder.embed(
+            mx.array([self._prompt_ids])
+        )[0]  # [prefix_len, dim]
+        mx.eval(self._prompt_embeds)
+
+        self._eos_id = self._tokenizer.eos_id
+        self._secs_per_token = SAMPLES_PER_TOKEN / self.SAMPLING_RATE
+        # The streaming model has an inherent delay: text for audio at position P
+        # is generated at decoder position P + n_delay. Compensate timestamps.
+        self._delay_secs = self._n_delay * self._secs_per_token
+
+        self._reset_state()
+
+    # -- state management --
+
+    def _reset_state(self):
+        """Reset all incremental state for a fresh utterance."""
+        # Audio accumulation
+        self._pending = np.zeros(0, dtype=np.float32)
+        # Mel overlap
+        self._mel_overlap: np.ndarray | None = None
+        # Encoder incremental state
+        self._conv_tail1 = None
+        self._conv_tail2 = None
+        self._enc_cache = None
+        self._ds_remainder = None
+        # Audio embeddings not yet decoded
+        self._audio_embeds: mx.array | None = None
+        # Decoder state
+        self._dec_cache: list[SlidingKVCache] | None = None
+        self._last_token: mx.array | None = None
+        # Bookkeeping
+        self._samples_encoded = 0
+        self._positions_decoded = 0
+        self._prefilled = False
+        self._first_chunk = True
+        # Text state
+        self._full_text = ""
+        self._n_text_tokens = 0
+        self._n_committed_words = 0
+        self._time_offset = 0.0
+        # Per-word audio position tracking: decoder position (relative to prefix)
+        # where each word in _full_text started and ended
+        self._word_audio_starts: list[int] = []   # audio pos where word i started
+        self._word_audio_ends: list[int] = []     # audio pos where word i last produced a token
+        self._current_word_pos: Optional[int] = None  # audio pos of current (incomplete) word's first token
+
+    # -- audio ingestion --
+
+    def insert_audio_chunk(self, audio: np.ndarray, audio_stream_end_time: float):
+        self.end = audio_stream_end_time
+        self._pending = np.append(self._pending, audio)
+        self.audio_buffer = self._pending
+
+    # -- core processing --
+
+    def process_iter(self, is_last=False) -> Tuple[List[ASRToken], float]:
+        try:
+            return self._step(is_last)
+        except Exception as e:
+            logger.warning("[voxtral-mlx] process_iter error: %s", e, exc_info=True)
+            return [], self.end
+
+    def _step(self, is_last: bool) -> Tuple[List[ASRToken], float]:
+        # 1. Encode any new audio
+        self._encode_pending()
+
+        if self._audio_embeds is None:
+            return [], self.end
+
+        # 2. Compute how many positions we can safely decode
+        total_safe = LEFT_PAD_TOKENS + self._samples_encoded // SAMPLES_PER_TOKEN
+        n_available = self._audio_embeds.shape[0]
+        n_decodable = min(n_available, total_safe - self._positions_decoded)
+
+        if n_decodable <= 0:
+            return [], self.end
+
+        # 3. Prefill if needed
+        if not self._prefilled:
+            if self._positions_decoded + n_available < self._prefix_len:
+                return [], self.end
+            self._do_prefill()
+            # Re-check after consuming prefix embeddings
+            n_available = self._audio_embeds.shape[0] if self._audio_embeds is not None else 0
+            n_decodable = min(n_available, total_safe - self._positions_decoded)
+
+        if n_decodable <= 0 or self._audio_embeds is None:
+            return [], self.end
+
+        # 4. Decode available positions
+        hit_eos = self._decode_positions(n_decodable)
+
+        if hit_eos:
+            # Flush words, reset for next utterance
+            words = self._flush_all_words()
+            logger.debug(
+                "[voxtral-mlx] EOS hit during stream: flushed %d words, "
+                "samples_encoded=%d (%.2fs), text='%s'",
+                len(words), self._samples_encoded,
+                self._samples_encoded / self.SAMPLING_RATE,
+                self._full_text[-60:] if self._full_text else "",
+            )
+            saved_offset = self._time_offset
+            self._reset_state()
+            self._time_offset = saved_offset
+            return words, self.end
+
+        # 5. Extract committed words (all but the last, which may still grow)
+        return self._extract_committed_words(), self.end
+
+    def _encode_pending(self):
+        """Feed pending audio through the incremental encoder."""
+        available = len(self._pending)
+        if available < SAMPLES_PER_TOKEN:
+            return
+
+        if self._first_chunk:
+            # First chunk: prepend silence for left-padding
+            n_take = (available // SAMPLES_PER_TOKEN) * SAMPLES_PER_TOKEN
+            left_pad = np.zeros(LEFT_PAD_TOKENS * SAMPLES_PER_TOKEN, dtype=np.float32)
+            chunk = np.concatenate([left_pad, self._pending[:n_take]])
+            self._pending = self._pending[n_take:]
+            self._samples_encoded += n_take
+            self._first_chunk = False
+        else:
+            n_take = (available // SAMPLES_PER_TOKEN) * SAMPLES_PER_TOKEN
+            chunk = self._pending[:n_take]
+            self._pending = self._pending[n_take:]
+            self._samples_encoded += n_take
+
+        mel, self._mel_overlap = compute_mel_streaming(chunk, self._mel_overlap)
+
+        embeds, self._conv_tail1, self._conv_tail2, self._enc_cache, self._ds_remainder = (
+            self._model.encode_incremental(
+                mel, self._conv_tail1, self._conv_tail2, self._enc_cache, self._ds_remainder
+            )
+        )
+
+        if embeds is not None:
+            mx.eval(embeds)
+            if self._audio_embeds is not None:
+                self._audio_embeds = mx.concatenate([self._audio_embeds, embeds])
+            else:
+                self._audio_embeds = embeds
+
+        self.audio_buffer = self._pending
+
+    def _do_prefill(self):
+        """Run the decoder prefill pass over the prompt + first audio embeddings."""
+        n_dec_layers = len(self._model.decoder.blocks)
+        self._dec_cache = [SlidingKVCache(_DECODER_WINDOW) for _ in range(n_dec_layers)]
+
+        prefix_embeds = self._prompt_embeds + self._audio_embeds[: self._prefix_len]
+        prefix_embeds = prefix_embeds[None, :, :]  # [1, prefix_len, dim]
+
+        logits = self._model.decode(prefix_embeds, self._delay_cond, "causal", self._dec_cache)
+        mx.eval(logits, *[x for c in self._dec_cache for x in (c.keys, c.values)])
+
+        self._last_token = self._sample(logits)
+        mx.async_eval(self._last_token)
+
+        # Remove consumed prefix embeddings
+        self._audio_embeds = self._audio_embeds[self._prefix_len :]
+        if self._audio_embeds.shape[0] == 0:
+            self._audio_embeds = None
+        self._positions_decoded = self._prefix_len
+        self._prefilled = True
+
+    def _decode_positions(self, n: int) -> bool:
+        """Autoregressively decode *n* positions.  Returns True on EOS."""
+        base_pos = self._positions_decoded  # absolute position before this batch
+        for i in range(n):
+            tok_embed = self._model.decoder.embed(self._last_token.reshape(1, 1))[0, 0]
+            combined = (self._audio_embeds[i] + tok_embed)[None, None, :]
+            logits = self._model.decode(combined, self._delay_cond, mask=None, cache=self._dec_cache)
+            next_tok = self._sample(logits)
+            mx.async_eval(next_tok)
+
+            token_id = self._last_token.item()
+            if token_id == self._eos_id:
+                # Close the current word if one is being built
+                if self._current_word_pos is not None:
+                    self._word_audio_ends.append(base_pos + i - self._prefix_len)
+                    self._current_word_pos = None
+                self._trim_embeds(i)
+                self._positions_decoded += i
+                return True
+
+            text = self._tokenizer.decode(
+                [token_id], special_token_policy=SpecialTokenPolicy.IGNORE
+            )
+
+            if text:
+                audio_pos = base_pos + i - self._prefix_len
+
+                # Detect word boundary: new word starts with space or is the very first text
+                if text.lstrip() != text or not self._full_text:
+                    # Close previous word if exists
+                    if self._current_word_pos is not None:
+                        self._word_audio_ends.append(audio_pos)
+                    # Start new word
+                    self._word_audio_starts.append(audio_pos)
+                    self._current_word_pos = audio_pos
+                elif self._current_word_pos is None:
+                    # First token of first word (no leading space)
+                    self._word_audio_starts.append(audio_pos)
+                    self._current_word_pos = audio_pos
+
+                self._full_text += text
+                self._n_text_tokens += 1
+
+            if i > 0 and i % 256 == 0:
+                mx.clear_cache()
+
+            self._last_token = next_tok
+
+        self._positions_decoded += n
+        self._trim_embeds(n)
+        return False
+
+    def _trim_embeds(self, n_consumed: int):
+        if self._audio_embeds is not None and self._audio_embeds.shape[0] > n_consumed:
+            self._audio_embeds = self._audio_embeds[n_consumed:]
+        else:
+            self._audio_embeds = None
+
+    def _sample(self, logits: mx.array) -> mx.array:
+        return mx.argmax(logits[0, -1:], axis=-1).squeeze()
+
+    # -- word extraction --
+
+    def _audio_pos_to_time(self, pos: int) -> float:
+        """Convert an audio position (relative to prefix end) to seconds."""
+        return max(0.0, pos * self._secs_per_token - self._delay_secs + self._time_offset)
+
+    def _word_time_range(self, word_idx: int, n_words: int) -> Tuple[float, float]:
+        """Compute (start, end) time for a word using tracked word positions."""
+        starts = self._word_audio_starts
+        ends = self._word_audio_ends
+
+        if not starts:
+            return self._time_offset, self._time_offset
+
+        # Get start position for this word
+        if word_idx < len(starts):
+            t0 = self._audio_pos_to_time(starts[word_idx])
+        else:
+            # Fallback: estimate from last known position
+            last_pos = ends[-1] if ends else starts[-1]
+            t0 = self._audio_pos_to_time(last_pos + 1)
+
+        # Get end position: use the start of the next word, or the end of this word
+        if word_idx + 1 < len(starts):
+            t1 = self._audio_pos_to_time(starts[word_idx + 1])
+        elif word_idx < len(ends):
+            t1 = self._audio_pos_to_time(ends[word_idx] + 1)
+        else:
+            # Last word, still being built: use last known position + 1 token
+            last_pos = starts[word_idx] if word_idx < len(starts) else (ends[-1] if ends else 0)
+            t1 = self._audio_pos_to_time(last_pos + 1)
+
+        return t0, t1
+
+    def _extract_committed_words(self) -> List[ASRToken]:
+        """Return complete words (all except the last which may still grow)."""
+        if not self._full_text:
+            return []
+        words = self._full_text.split()
+        tokens: List[ASRToken] = []
+        n_total = max(len(words), 1)
+
+        while len(words) > self._n_committed_words + 1:
+            w = words[self._n_committed_words]
+            idx = self._n_committed_words
+            t0, t1 = self._word_time_range(idx, n_total)
+            label = w if idx == 0 else " " + w
+            tokens.append(ASRToken(start=t0, end=t1, text=label))
+            self._n_committed_words += 1
+
+        return tokens
+
+    def _flush_all_words(self) -> List[ASRToken]:
+        """Flush every word including the last partial one."""
+        if not self._full_text:
+            return []
+        words = self._full_text.split()
+        tokens: List[ASRToken] = []
+        n_total = max(len(words), 1)
+
+        while self._n_committed_words < len(words):
+            w = words[self._n_committed_words]
+            idx = self._n_committed_words
+            t0, t1 = self._word_time_range(idx, n_total)
+            label = w if idx == 0 else " " + w
+            tokens.append(ASRToken(start=t0, end=t1, text=label))
+            self._n_committed_words += 1
+
+        return tokens
+
+    # -- interface methods --
+
+    def get_buffer(self) -> Transcript:
+        if not self._full_text:
+            return Transcript(start=None, end=None, text="")
+        words = self._full_text.split()
+        remaining = words[self._n_committed_words :]
+        if remaining:
+            return Transcript(start=self.end, end=self.end, text=" ".join(remaining))
+        return Transcript(start=None, end=None, text="")
+
+    def start_silence(self) -> Tuple[List[ASRToken], float]:
+        words = self._flush_all_words()
+        logger.info("[voxtral-mlx] start_silence: flushed %d words", len(words))
+        return words, self.end
+
+    def end_silence(self, silence_duration: float, offset: float):
+        self._time_offset += silence_duration
+        self.end += silence_duration
+
+    def new_speaker(self, change_speaker):
+        self.start_silence()
+
+    def warmup(self, audio, init_prompt=""):
+        pass
+
+    def finish(self) -> Tuple[List[ASRToken], float]:
+        logger.debug(
+            "[voxtral-mlx] finish: pending=%d samples, audio_embeds=%s, "
+            "samples_encoded=%d, positions_decoded=%d, prefilled=%s, text so far='%s'",
+            len(self._pending),
+            self._audio_embeds.shape if self._audio_embeds is not None else None,
+            self._samples_encoded,
+            self._positions_decoded,
+            self._prefilled,
+            self._full_text[-80:] if self._full_text else "",
+        )
+
+        # Align pending audio to SAMPLES_PER_TOKEN boundary so nothing is lost
+        remainder = len(self._pending) % SAMPLES_PER_TOKEN
+        if remainder > 0:
+            align_pad = SAMPLES_PER_TOKEN - remainder
+        else:
+            align_pad = 0
+
+        # Add alignment + right-padding silence
+        total_pad = align_pad + RIGHT_PAD_TOKENS * SAMPLES_PER_TOKEN
+        if total_pad > 0:
+            self._pending = np.append(
+                self._pending, np.zeros(total_pad, dtype=np.float32)
+            )
+
+        # Encode remaining audio (including right-padding)
+        self._encode_pending()
+
+        logger.debug(
+            "[voxtral-mlx] finish after encode: audio_embeds=%s, pending=%d",
+            self._audio_embeds.shape if self._audio_embeds is not None else None,
+            len(self._pending),
+        )
+
+        hit_eos = False
+
+        # Decode everything that's left from right-padding
+        if self._audio_embeds is not None and self._prefilled:
+            hit_eos = self._decode_positions(self._audio_embeds.shape[0])
+            logger.debug(
+                "[voxtral-mlx] finish decode: hit_eos=%s, text='%s'",
+                hit_eos, self._full_text[-80:] if self._full_text else "",
+            )
+
+        # Flush last token if it wasn't EOS
+        if self._last_token is not None:
+            tid = self._last_token.item()
+            if tid != self._eos_id:
+                text = self._tokenizer.decode(
+                    [tid], special_token_policy=SpecialTokenPolicy.IGNORE
+                )
+                if text:
+                    last_pos = self._positions_decoded - self._prefix_len
+                    # Check if this starts a new word
+                    if text.lstrip() != text or not self._full_text:
+                        if self._current_word_pos is not None:
+                            self._word_audio_ends.append(last_pos)
+                        self._word_audio_starts.append(last_pos)
+                        self._current_word_pos = last_pos
+                    elif self._current_word_pos is None:
+                        self._word_audio_starts.append(last_pos)
+                        self._current_word_pos = last_pos
+                    self._full_text += text
+                    self._n_text_tokens += 1
+
+        # Close the last word if still open
+        if self._current_word_pos is not None:
+            last_pos = self._positions_decoded - self._prefix_len
+            self._word_audio_ends.append(last_pos)
+            self._current_word_pos = None
+
+        words = self._flush_all_words()
+        logger.info("[voxtral-mlx] finish: flushed %d words", len(words))
+        return words, self.end

whisperlivekit/warmup.py

@@ -7,6 +7,7 @@ def load_file(warmup_file=None, timeout=5):
     import os
     import tempfile
     import urllib.request
+
     import librosa
 
     if warmup_file == "":

whisperlivekit/web/web_interface.py

@@ -1,6 +1,6 @@
-import logging
-import importlib.resources as resources
 import base64
+import importlib.resources as resources
+import logging
 
 logger = logging.getLogger(__name__)
 
@@ -96,11 +96,13 @@ def get_inline_ui_html():
 
 if __name__ == '__main__':
     
+    import pathlib
+
+    import uvicorn
     from fastapi import FastAPI
     from fastapi.responses import HTMLResponse
-    import uvicorn
     from starlette.staticfiles import StaticFiles
-    import pathlib
+
     import whisperlivekit.web as webpkg
     
     app = FastAPI()    

whisperlivekit/whisper/__init__.py

@@ -4,15 +4,17 @@ import json
 import os
 import urllib
 import warnings
+from pathlib import Path
 from typing import Dict, List, Optional, Union
 
 import torch
-from tqdm import tqdm
-from pathlib import Path
 from torch import Tensor
+from tqdm import tqdm
 
-from whisperlivekit.whisper.audio import load_audio, log_mel_spectrogram, pad_or_trim
-from whisperlivekit.whisper.decoding import DecodingOptions, DecodingResult, decode, detect_language
+from whisperlivekit.whisper.audio import (load_audio, log_mel_spectrogram,
+                                          pad_or_trim)
+from whisperlivekit.whisper.decoding import (DecodingOptions, DecodingResult,
+                                             decode, detect_language)
 from whisperlivekit.whisper.model import ModelDimensions, Whisper
 from whisperlivekit.whisper.transcribe import transcribe
 from whisperlivekit.whisper.version import __version__
@@ -106,7 +108,7 @@ def available_models() -> List[str]:
 def _infer_dims_from_config(path: str) -> Optional[ModelDimensions]:
     """
     attempt to infer ModelDimensions from a HF style config.json located
-    next to the given checkpoint, usefull for distilled models
+    next to the given checkpoint, usefull for distilled models/MLX models.
     """
     candidates = []
     if os.path.isdir(path):
@@ -120,6 +122,25 @@ def _infer_dims_from_config(path: str) -> Optional[ModelDimensions]:
         with open(candidate, "r", encoding="utf-8") as f:
             config = json.load(f)
 
+        # native Whisper format
+        native_keys = ["n_mels", "n_audio_ctx", "n_audio_state", "n_audio_head",
+                       "n_audio_layer", "n_vocab", "n_text_ctx", "n_text_state",
+                       "n_text_head", "n_text_layer"]
+        if all(k in config for k in native_keys):
+            return ModelDimensions(
+                n_mels=config["n_mels"],
+                n_audio_ctx=config["n_audio_ctx"],
+                n_audio_state=config["n_audio_state"],
+                n_audio_head=config["n_audio_head"],
+                n_audio_layer=config["n_audio_layer"],
+                n_vocab=config["n_vocab"],
+                n_text_ctx=config["n_text_ctx"],
+                n_text_state=config["n_text_state"],
+                n_text_head=config["n_text_head"],
+                n_text_layer=config["n_text_layer"],
+            )
+
+        # HuggingFace format
         try:
             return ModelDimensions(
                 n_mels=config["num_mel_bins"],
@@ -234,6 +255,24 @@ def _convert_hf_state_dict(state_dict: Dict[str, torch.Tensor]) -> Dict[str, tor
     return converted if converted else state_dict
 
 
+def _convert_mlx_state_dict(state_dict: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
+    """
+    Converts an mlx whisper checkpoint to a default openai whisper one
+    """
+    if not any("mlp1" in k or "mlp2" in k for k in state_dict):
+        return state_dict
+
+    converted = {}
+    for key, value in state_dict.items():
+        if key == "alignment_heads":
+            continue
+        
+        new_key = key.replace(".mlp1.", ".mlp.0.").replace(".mlp2.", ".mlp.2.")
+        converted[new_key] = value
+
+    return converted
+
+
 def _load_lora_state(lora_path: str):
     safe_path = os.path.join(lora_path, "adapter_model.safetensors")
     bin_path = os.path.join(lora_path, "adapter_model.bin")
@@ -262,9 +301,49 @@ def _collapse_hf_module_name(module: str):
     return module
 
 
+def _resolve_lora_path(lora_path: Optional[str]) -> Optional[str]:
+    """
+    Resolve LoRA adapter path - handles both local paths and HuggingFace repo IDs.
+    
+    If lora_path is a local directory containing adapter files, returns it as-is.
+    If lora_path looks like a HuggingFace repo ID (contains '/'), downloads and caches it.
+    """
+    if not lora_path:
+        return None
+    
+    # Check if it's already a valid local path
+    if os.path.isdir(lora_path):
+        config_path = os.path.join(lora_path, "adapter_config.json")
+        if os.path.isfile(config_path):
+            return lora_path
+    
+    # Try to download from HuggingFace Hub
+    if "/" in lora_path:
+        try:
+            from huggingface_hub import snapshot_download
+            local_path = snapshot_download(
+                repo_id=lora_path,
+                allow_patterns=["adapter_config.json", "adapter_model.*"],
+            )
+            return local_path
+        except Exception as e:
+            raise FileNotFoundError(
+                f"Could not find LoRA adapter at local path or HuggingFace Hub: {lora_path}. Error: {e}"
+            )
+    
+    raise FileNotFoundError(
+        f"LoRA path '{lora_path}' is not a valid local directory or HuggingFace repo ID."
+    )
+
+
 def _apply_lora_adapter(state_dict: Dict[str, Tensor], lora_path: Optional[str]):
     if not lora_path:
         return
+    
+    # Resolve path (handles HuggingFace Hub download)
+    lora_path = _resolve_lora_path(lora_path)
+    if not lora_path:
+        return
 
     config_path = os.path.join(lora_path, "adapter_config.json")
     if not os.path.isfile(config_path):
@@ -317,6 +396,75 @@ def _apply_lora_adapter(state_dict: Dict[str, Tensor], lora_path: Optional[str])
         )
 
 
+def _load_checkpoint(
+    file_path: Union[str, Path],
+    device: str,
+    in_memory: bool = False,
+    checkpoint_bytes: Optional[bytes] = None,
+) -> Dict[str, torch.Tensor]:
+    """
+    Load a checkpoint from a single file.
+    
+    Handles .pt, .bin, and .safetensors formats.
+    """
+    if checkpoint_bytes is not None:
+        with io.BytesIO(checkpoint_bytes) as fp:
+            return torch.load(fp, map_location=device)
+    
+    file_path = Path(file_path)
+    suffix = file_path.suffix.lower()
+    
+    if suffix == '.safetensors':
+        try:
+            from safetensors.torch import load_file
+        except ImportError:
+            raise ImportError(
+                "Please install safetensors to load .safetensors model files: `pip install safetensors`"
+            )
+        return load_file(str(file_path), device=device)
+    else:
+        if in_memory:
+            with open(file_path, "rb") as f:
+                checkpoint_bytes = f.read()
+            with io.BytesIO(checkpoint_bytes) as fp:
+                return torch.load(fp, map_location=device)
+        else:
+            with open(file_path, "rb") as fp:
+                return torch.load(fp, map_location=device)
+
+
+def _load_sharded_checkpoint(
+    shard_files: List[Path],
+    device: str,
+) -> Dict[str, torch.Tensor]:
+    """
+    Load a sharded checkpoint (multiple .safetensors or .bin files).
+    
+    Merges all shards into a single state dict.
+    """
+    merged_state_dict = {}
+    first_suffix = shard_files[0].suffix.lower()
+    
+    if first_suffix == '.safetensors':
+        try:
+            from safetensors.torch import load_file
+        except ImportError:
+            raise ImportError(
+                "Please install safetensors to load sharded .safetensors model: `pip install safetensors`"
+            )
+        for shard_path in shard_files:
+            shard_dict = load_file(str(shard_path), device=device)
+            merged_state_dict.update(shard_dict)
+    else:
+        for shard_path in shard_files:
+            with open(shard_path, "rb") as fp:
+                shard_dict = torch.load(fp, map_location=device)
+            if isinstance(shard_dict, dict):
+                merged_state_dict.update(shard_dict)
+    
+    return merged_state_dict
+
+
 def load_model(
     name: str,
     device: Optional[Union[str, torch.device]] = None,
@@ -334,6 +482,8 @@ def load_model(
     name : str
         one of the official model names listed by `whisper.available_models()`, or
         path to a model checkpoint containing the model dimensions and the model state_dict.
+        Can be a single file (.pt, .bin, .safetensors), a directory containing model files,
+        or a sharded model directory with files like model-00001-of-00002.safetensors.
     device : Union[str, torch.device]
         the PyTorch device to put the model into
     download_root: str
@@ -348,16 +498,51 @@ def load_model(
     model : Whisper
         The Whisper ASR model instance
     """
+    from whisperlivekit.model_paths import detect_model_format
 
     if device is None:
         device = "cuda" if torch.cuda.is_available() else "cpu"
     if download_root is None:
         default = os.path.join(os.path.expanduser("~"), ".cache")
         download_root = os.path.join(os.getenv("XDG_CACHE_HOME", default), "whisper")
+    
+    checkpoint = None
+    model_path_for_config = name  # Used to find config.json for dims inference
+    
     if name in _MODELS:
-        checkpoint_file = _download(_MODELS[name], download_root, in_memory)        
+        checkpoint_file = _download(_MODELS[name], download_root, in_memory)
+        if in_memory:
+            checkpoint = _load_checkpoint(None, device, checkpoint_bytes=checkpoint_file)
+        else:
+            checkpoint = _load_checkpoint(checkpoint_file, device)
     elif os.path.isfile(name):
-        checkpoint_file = open(name, "rb").read() if in_memory else name
+        if in_memory:
+            with open(name, "rb") as f:
+                checkpoint_bytes = f.read()
+            checkpoint = _load_checkpoint(None, device, checkpoint_bytes=checkpoint_bytes)
+        else:
+            checkpoint = _load_checkpoint(name, device)
+        model_path_for_config = name
+    elif os.path.isdir(name):
+        model_info = detect_model_format(name)
+        
+        if not model_info.has_pytorch:
+            raise RuntimeError(
+                f"No PyTorch checkpoint found in directory {name}. "
+                f"Expected .pt, .bin, or .safetensors file(s)."
+            )
+        
+        if model_info.is_sharded:
+            checkpoint = _load_sharded_checkpoint(model_info.pytorch_files, device)
+        else:
+            single_file = model_info.pytorch_files[0]
+            if in_memory:
+                with open(single_file, "rb") as f:
+                    checkpoint_bytes = f.read()
+                checkpoint = _load_checkpoint(None, device, checkpoint_bytes=checkpoint_bytes)
+            else:
+                checkpoint = _load_checkpoint(single_file, device)
+        model_path_for_config = name
     else:
         raise RuntimeError(
             f"Model {name} not found; available models = {available_models()}"
@@ -367,34 +552,23 @@ def load_model(
     if custom_alignment_heads:
         alignment_heads = custom_alignment_heads.encode()
 
-    if isinstance(checkpoint_file, Path) and checkpoint_file.suffix == '.safetensors':
-        try:
-            from safetensors.torch import load_file
-        except ImportError:
-            raise ImportError("Please install safetensors to load .safetensors model files: `pip install safetensors`")
-        if in_memory:
-            checkpoint = load_file(checkpoint_file, device=device)
-        else:
-            checkpoint = load_file(checkpoint_file, device=device)
-    else:
-        with (
-            io.BytesIO(checkpoint_file) if in_memory else open(checkpoint_file, "rb")
-        ) as fp:
-            checkpoint = torch.load(fp, map_location=device)
-    del checkpoint_file
-
     dims_cfg = checkpoint.get("dims") if isinstance(checkpoint, dict) else None
     if isinstance(checkpoint, dict) and "model_state_dict" in checkpoint:
         state_dict = checkpoint["model_state_dict"]
     else:
         state_dict = checkpoint
+    
+    if alignment_heads is None and "alignment_heads" in state_dict:
+        alignment_heads = state_dict["alignment_heads"]
+    
     state_dict = _convert_hf_state_dict(state_dict)
+    state_dict = _convert_mlx_state_dict(state_dict)
     _apply_lora_adapter(state_dict, lora_path)
 
     if dims_cfg is not None:
         dims = ModelDimensions(**dims_cfg)
     else:
-        dims = _infer_dims_from_config(name)
+        dims = _infer_dims_from_config(model_path_for_config)
         if dims is None:
             raise RuntimeError(
                 "Could not determine model dimensions. "
@@ -414,8 +588,13 @@ def load_model(
     model.load_state_dict(state_dict)
 
     if alignment_heads is not None:
-        model.set_alignment_heads(alignment_heads)
-
+        if isinstance(alignment_heads, bytes):
+            model.set_alignment_heads(alignment_heads)
+        elif isinstance(alignment_heads, torch.Tensor): #for mlx whisper
+            mask = torch.zeros(dims.n_text_layer, dims.n_text_head, dtype=torch.bool)
+            for layer, head in alignment_heads.tolist():
+                mask[layer, head] = True
+            model.register_buffer("alignment_heads", mask.to_sparse(), persistent=False)
     return model.to(device)
 
 

whisperlivekit/whisper/decoding.py

@@ -1,5 +1,6 @@
 from dataclasses import dataclass, field, replace
-from typing import TYPE_CHECKING, Dict, Iterable, List, Optional, Sequence, Tuple, Union
+from typing import (TYPE_CHECKING, Dict, Iterable, List, Optional, Sequence,
+                    Tuple, Union)
 
 import numpy as np
 import torch
@@ -146,16 +147,13 @@ class PyTorchInference(Inference):
         self.model: "Whisper" = model
         self.initial_token_length = initial_token_length
         self.kv_cache = {}
-        self.hooks = []
 
-        key_modules = [block.attn.key for block in self.model.decoder.blocks]
-        value_modules = [block.attn.value for block in self.model.decoder.blocks]
-        self.kv_modules = key_modules + value_modules
+        self.kv_cache_ids = []
+        for block in self.model.decoder.blocks:
+            self.kv_cache_ids.append(block.attn.key_cache_id)
+            self.kv_cache_ids.append(block.attn.value_cache_id)
 
     def logits(self, tokens: Tensor, audio_features: Tensor) -> Tensor:
-        if not self.kv_cache:
-            self.kv_cache, self.hooks = self.model.install_kv_cache_hooks()
-
         if tokens.shape[-1] > self.initial_token_length:
             # only need to use the last token except in the first forward pass
             tokens = tokens[:, -1:]
@@ -163,17 +161,14 @@ class PyTorchInference(Inference):
         return self.model.decoder(tokens, audio_features, kv_cache=self.kv_cache)
 
     def cleanup_caching(self):
-        for hook in self.hooks:
-            hook.remove()
-
         self.kv_cache = {}
-        self.hooks = []
 
     def rearrange_kv_cache(self, source_indices):
         if source_indices != list(range(len(source_indices))):
-            for module in self.kv_modules:
-                # update the key/value cache to contain the selected sequences
-                self.kv_cache[module] = self.kv_cache[module][source_indices].detach()
+            for cache_id in self.kv_cache_ids:
+                if cache_id in self.kv_cache:
+                    # update the key/value cache to contain the selected sequences
+                    self.kv_cache[cache_id] = self.kv_cache[cache_id][source_indices].detach()
 
 
 class SequenceRanker:

whisperlivekit/whisper/model.py

@@ -79,18 +79,23 @@ def disable_sdpa():
 
 
 class MultiHeadAttention(nn.Module):
-    use_sdpa = False  # Disable SDPA to ensure qk is always computed for hooks
+    use_sdpa = False  # Disable SDPA to ensure qk is always computed when needed
 
-    def __init__(self, n_state: int, n_head: int, cache_id: str = ""):
+    def __init__(self, n_state: int, n_head: int, cache_id: str = "", n_text_ctx: int = 448):
         super().__init__()
         self.n_head = n_head
+        self.n_text_ctx = n_text_ctx
         self.query = Linear(n_state, n_state)
         self.key = Linear(n_state, n_state, bias=False)
         self.value = Linear(n_state, n_state)
         self.out = Linear(n_state, n_state)
         self.cache_id = cache_id
-        self.key.cache_id = f"{cache_id}_key"
-        self.value.cache_id = f"{cache_id}_value"
+        # Cache IDs for key and value (used with dict-based kv_cache)
+        self.key_cache_id = f"{cache_id}_key"
+        self.value_cache_id = f"{cache_id}_value"
+        # Keep these for backward compatibility with hook-based caching
+        self.key.cache_id = self.key_cache_id
+        self.value.cache_id = self.value_cache_id
 
     def forward(
         self,
@@ -101,19 +106,45 @@ class MultiHeadAttention(nn.Module):
     ):
         q = self.query(x)
 
-        if kv_cache is None or xa is None or self.key not in kv_cache:
-            # hooks, if installed (i.e. kv_cache is not None), will prepend the cached kv tensors;
-            # otherwise, perform key/value projections for self- or cross-attention as usual.
-            k = self.key(x if xa is None else xa)
-            v = self.value(x if xa is None else xa)
+        if xa is None:
+            # Self-attention
+            k = self.key(x)
+            v = self.value(x)
+            if kv_cache is not None:
+                k, v = self._update_self_attn_cache(k, v, kv_cache)
         else:
-            # for cross-attention, calculate keys and values once and reuse in subsequent calls.
-            k = kv_cache[self.key]
-            v = kv_cache[self.value]
+            # Cross-attention: compute once and cache, or reuse from cache
+            if kv_cache is not None and self.key_cache_id in kv_cache:
+                k = kv_cache[self.key_cache_id]
+                v = kv_cache[self.value_cache_id]
+            else:
+                k = self.key(xa)
+                v = self.value(xa)
+                if kv_cache is not None:
+                    kv_cache[self.key_cache_id] = k
+                    kv_cache[self.value_cache_id] = v
 
         wv, qk = self.qkv_attention(q, k, v, mask)
         return self.out(wv), qk
 
+    def _update_self_attn_cache(
+        self, k: Tensor, v: Tensor, kv_cache: dict
+    ) -> Tuple[Tensor, Tensor]:
+        """Update self-attention kv cache by concatenating new k,v with cached values."""
+        if self.key_cache_id not in kv_cache or k.shape[1] > self.n_text_ctx:
+            # First token or context overflow: save as-is
+            kv_cache[self.key_cache_id] = k.detach()
+            kv_cache[self.value_cache_id] = v.detach()
+        else:
+            # Concatenate with existing cache
+            cached_k = kv_cache[self.key_cache_id]
+            cached_v = kv_cache[self.value_cache_id]
+            k = torch.cat([cached_k, k], dim=1).detach()
+            v = torch.cat([cached_v, v], dim=1).detach()
+            kv_cache[self.key_cache_id] = k
+            kv_cache[self.value_cache_id] = v
+        return k, v
+
     def qkv_attention(
         self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor] = None
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
@@ -143,14 +174,21 @@ class MultiHeadAttention(nn.Module):
 
 
 class ResidualAttentionBlock(nn.Module):
-    def __init__(self, n_state: int, n_head: int, cross_attention: bool = False, cache_id: str = ""):
+    def __init__(
+        self, n_state: int, n_head: int, cross_attention: bool = False, 
+        cache_id: str = "", n_text_ctx: int = 448
+    ):
         super().__init__()
 
-        self.attn = MultiHeadAttention(n_state, n_head, cache_id=f"{cache_id}_self_attn")
+        self.attn = MultiHeadAttention(
+            n_state, n_head, cache_id=f"{cache_id}_self_attn", n_text_ctx=n_text_ctx
+        )
         self.attn_ln = LayerNorm(n_state)
 
         self.cross_attn = (
-            MultiHeadAttention(n_state, n_head, cache_id=f"{cache_id}_cross_attn") if cross_attention else None
+            MultiHeadAttention(
+                n_state, n_head, cache_id=f"{cache_id}_cross_attn", n_text_ctx=n_text_ctx
+            ) if cross_attention else None
         )
         self.cross_attn_ln = LayerNorm(n_state) if cross_attention else None
 
@@ -166,12 +204,21 @@ class ResidualAttentionBlock(nn.Module):
         xa: Optional[Tensor] = None,
         mask: Optional[Tensor] = None,
         kv_cache: Optional[dict] = None,
-    ):
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        """
+        Returns:
+            x: The output tensor
+            cross_attn_qk: Cross-attention weights (if cross_attn exists), else None
+        """
         x = x + self.attn(self.attn_ln(x), mask=mask, kv_cache=kv_cache)[0]
+        cross_attn_qk = None
         if self.cross_attn:
-            x = x + self.cross_attn(self.cross_attn_ln(x), xa, kv_cache=kv_cache)[0]
+            cross_out, cross_attn_qk = self.cross_attn(
+                self.cross_attn_ln(x), xa, kv_cache=kv_cache
+            )
+            x = x + cross_out
         x = x + self.mlp(self.mlp_ln(x))
-        return x
+        return x, cross_attn_qk
 
 
 class AudioEncoder(nn.Module):
@@ -201,7 +248,7 @@ class AudioEncoder(nn.Module):
         x = (x + self.positional_embedding).to(x.dtype)
 
         for block in self.blocks:
-            x = block(x)
+            x, _ = block(x)  # Encoder blocks don't have cross-attention
 
         x = self.ln_post(x)
         return x
@@ -212,13 +259,17 @@ class TextDecoder(nn.Module):
         self, n_vocab: int, n_ctx: int, n_state: int, n_head: int, n_layer: int
     ):
         super().__init__()
+        self.n_ctx = n_ctx
 
         self.token_embedding = nn.Embedding(n_vocab, n_state)
         self.positional_embedding = nn.Parameter(torch.empty(n_ctx, n_state))
 
         self.blocks: Iterable[ResidualAttentionBlock] = nn.ModuleList(
             [
-                ResidualAttentionBlock(n_state, n_head, cross_attention=True, cache_id=f"dec_layer{i}")
+                ResidualAttentionBlock(
+                    n_state, n_head, cross_attention=True, 
+                    cache_id=f"dec_layer{i}", n_text_ctx=n_ctx
+                )
                 for i in range(n_layer)
             ]
         )
@@ -227,28 +278,57 @@ class TextDecoder(nn.Module):
         mask = torch.empty(n_ctx, n_ctx).fill_(-np.inf).triu_(1)
         self.register_buffer("mask", mask, persistent=False)
 
-    def forward(self, x: Tensor, xa: Tensor, kv_cache: Optional[dict] = None):
+    def forward(
+        self, 
+        x: Tensor, 
+        xa: Tensor, 
+        kv_cache: Optional[dict] = None,
+        return_cross_attn: bool = False,
+    ):
         """
         x : torch.LongTensor, shape = (batch_size, <= n_ctx)
             the text tokens
         xa : torch.Tensor, shape = (batch_size, n_audio_ctx, n_audio_state)
             the encoded audio features to be attended on
+        kv_cache : Optional[dict]
+            Dictionary to store/retrieve key-value cache for efficient decoding
+        return_cross_attn : bool
+            If True, return cross-attention weights from all decoder layers
+            
+        Returns
+        -------
+        logits : Tensor
+            The output logits
+        cross_attns : Optional[List[Tensor]]
+            List of cross-attention weights per layer (only if return_cross_attn=True)
         """
-        offset = next(iter(kv_cache.values())).shape[1] if kv_cache else 0
+        # Calculate offset from self-attention cache (not cross-attention which has audio length)
+        offset = 0
+        if kv_cache:
+            # Use the first decoder block's self-attention key cache to get token position
+            first_self_attn_key = self.blocks[0].attn.key_cache_id
+            if first_self_attn_key in kv_cache:
+                offset = kv_cache[first_self_attn_key].shape[1]
+        
         x = (
             self.token_embedding(x)
             + self.positional_embedding[offset : offset + x.shape[-1]]
         )
         x = x.to(xa.dtype)
 
+        cross_attns = [] if return_cross_attn else None
         for block in self.blocks:
-            x = block(x, xa, mask=self.mask, kv_cache=kv_cache)
+            x, cross_attn_qk = block(x, xa, mask=self.mask, kv_cache=kv_cache)
+            if return_cross_attn and cross_attn_qk is not None:
+                cross_attns.append(cross_attn_qk)
 
         x = self.ln(x)
         logits = (
             x @ torch.transpose(self.token_embedding.weight.to(x.dtype), 0, 1)
         ).float()
 
+        if return_cross_attn:
+            return logits, cross_attns
         return logits
 
 
@@ -292,8 +372,18 @@ class Whisper(nn.Module):
     def embed_audio(self, mel: torch.Tensor):
         return self.encoder(mel)
 
-    def logits(self, tokens: torch.Tensor, audio_features: torch.Tensor):
-        return self.decoder(tokens, audio_features)
+    def logits(
+        self, 
+        tokens: torch.Tensor, 
+        audio_features: torch.Tensor,
+        kv_cache: Optional[dict] = None,
+        return_cross_attn: bool = False,
+    ):
+        return self.decoder(
+            tokens, audio_features, 
+            kv_cache=kv_cache, 
+            return_cross_attn=return_cross_attn
+        )
 
     def forward(
         self, mel: torch.Tensor, tokens: torch.Tensor
@@ -312,39 +402,6 @@ class Whisper(nn.Module):
     def num_languages(self):
         return self.dims.n_vocab - 51765 - int(self.is_multilingual)
 
-    def install_kv_cache_hooks(self, cache: Optional[dict] = None):
-        """
-        The `MultiHeadAttention` module optionally accepts `kv_cache` which stores the key and value
-        tensors calculated for the previous positions. This method returns a dictionary that stores
-        all caches, and the necessary hooks for the key and value projection modules that save the
-        intermediate tensors to be reused during later calculations.
-
-        Returns
-        -------
-        cache : Dict[nn.Module, torch.Tensor]
-            A dictionary object mapping the key/value projection modules to its cache
-        hooks : List[RemovableHandle]
-            List of PyTorch RemovableHandle objects to stop the hooks to be called
-        """
-        cache = {**cache} if cache is not None else {}
-        hooks = []
-
-        def save_to_cache(module, _, output):
-            if module not in cache or output.shape[1] > self.dims.n_text_ctx:
-                # save as-is, for the first token or cross attention
-                cache[module] = output
-            else:
-                cache[module] = torch.cat([cache[module], output], dim=1).detach()
-            return cache[module]
-
-        def install_hooks(layer: nn.Module):
-            if isinstance(layer, MultiHeadAttention):
-                hooks.append(layer.key.register_forward_hook(save_to_cache))
-                hooks.append(layer.value.register_forward_hook(save_to_cache))
-
-        self.decoder.apply(install_hooks)
-        return cache, hooks
-
     detect_language = detect_language_function
     transcribe = transcribe_function
     decode = decode_function

whisperlivekit/whisper/tokenizer.py

@@ -296,10 +296,15 @@ class Tokenizer:
             current_tokens.append(token)
             decoded = self.decode_with_timestamps(current_tokens)
 
-            if (
-                replacement_char not in decoded
-                or decoded_full[unicode_offset + decoded.index(replacement_char)]
-                == replacement_char
+            try:
+                replacement_char_index = decoded.index(replacement_char)
+                replacement_char_index += unicode_offset
+            except ValueError:
+                replacement_char_index = None
+
+            if replacement_char_index is None or (
+                replacement_char_index < len(decoded_full)
+                and decoded_full[replacement_char_index] == replacement_char
             ):
                 words.append(decoded)
                 word_tokens.append(current_tokens)

whisperlivekit/whisper/transcribe.py

@@ -8,28 +8,13 @@ import numpy as np
 import torch
 import tqdm
 
-from .audio import (
-    FRAMES_PER_SECOND,
-    HOP_LENGTH,
-    N_FRAMES,
-    N_SAMPLES,
-    SAMPLE_RATE,
-    log_mel_spectrogram,
-    pad_or_trim,
-)
+from .audio import (FRAMES_PER_SECOND, HOP_LENGTH, N_FRAMES, N_SAMPLES,
+                    SAMPLE_RATE, log_mel_spectrogram, pad_or_trim)
 from .decoding import DecodingOptions, DecodingResult
 from .timing import add_word_timestamps
 from .tokenizer import LANGUAGES, TO_LANGUAGE_CODE, get_tokenizer
-from .utils import (
-    exact_div,
-    format_timestamp,
-    get_end,
-    get_writer,
-    make_safe,
-    optional_float,
-    optional_int,
-    str2bool,
-)
+from .utils import (exact_div, format_timestamp, get_end, get_writer,
+                    make_safe, optional_float, optional_int, str2bool)
 
 if TYPE_CHECKING:
     from .model import Whisper

whisperlivekit/whisper/val.py

@@ -0,0 +1,200 @@
+"""
+The most atomic way to train and inference a GPT in pure, dependency-free Python.
+This file is the complete algorithm.
+Everything else is just efficiency.
+
+@karpathy
+"""
+
+import os       # os.path.exists
+import math     # math.log, math.exp
+import random   # random.seed, random.choices, random.gauss, random.shuffle
+random.seed(42) # Let there be order among chaos
+
+# Let there be an input dataset `docs`: list[str] of documents (e.g. a dataset of names)
+if not os.path.exists('input.txt'):
+    import urllib.request
+    names_url = 'https://raw.githubusercontent.com/karpathy/makemore/refs/heads/master/names.txt'
+    urllib.request.urlretrieve(names_url, 'input.txt')
+docs = [l.strip() for l in open('input.txt').read().strip().split('\n') if l.strip()] # list[str] of documents
+random.shuffle(docs)
+print(f"num docs: {len(docs)}")
+
+# Let there be a Tokenizer to translate strings to discrete symbols and back
+uchars = sorted(set(''.join(docs))) # unique characters in the dataset become token ids 0..n-1
+BOS = len(uchars) # token id for the special Beginning of Sequence (BOS) token
+vocab_size = len(uchars) + 1 # total number of unique tokens, +1 is for BOS
+print(f"vocab size: {vocab_size}")
+
+# Let there be Autograd, to recursively apply the chain rule through a computation graph
+class Value:
+    __slots__ = ('data', 'grad', '_children', '_local_grads') # Python optimization for memory usage
+
+    def __init__(self, data, children=(), local_grads=()):
+        self.data = data                # scalar value of this node calculated during forward pass
+        self.grad = 0                   # derivative of the loss w.r.t. this node, calculated in backward pass
+        self._children = children       # children of this node in the computation graph
+        self._local_grads = local_grads # local derivative of this node w.r.t. its children
+
+    def __add__(self, other):
+        other = other if isinstance(other, Value) else Value(other)
+        return Value(self.data + other.data, (self, other), (1, 1))
+
+    def __mul__(self, other):
+        other = other if isinstance(other, Value) else Value(other)
+        return Value(self.data * other.data, (self, other), (other.data, self.data))
+
+    def __pow__(self, other): return Value(self.data**other, (self,), (other * self.data**(other-1),))
+    def log(self): return Value(math.log(self.data), (self,), (1/self.data,))
+    def exp(self): return Value(math.exp(self.data), (self,), (math.exp(self.data),))
+    def relu(self): return Value(max(0, self.data), (self,), (float(self.data > 0),))
+    def __neg__(self): return self * -1
+    def __radd__(self, other): return self + other
+    def __sub__(self, other): return self + (-other)
+    def __rsub__(self, other): return other + (-self)
+    def __rmul__(self, other): return self * other
+    def __truediv__(self, other): return self * other**-1
+    def __rtruediv__(self, other): return other * self**-1
+
+    def backward(self):
+        topo = []
+        visited = set()
+        def build_topo(v):
+            if v not in visited:
+                visited.add(v)
+                for child in v._children:
+                    build_topo(child)
+                topo.append(v)
+        build_topo(self)
+        self.grad = 1
+        for v in reversed(topo):
+            for child, local_grad in zip(v._children, v._local_grads):
+                child.grad += local_grad * v.grad
+
+# Initialize the parameters, to store the knowledge of the model.
+n_embd = 16     # embedding dimension
+n_head = 4      # number of attention heads
+n_layer = 1     # number of layers
+block_size = 16 # maximum sequence length
+head_dim = n_embd // n_head # dimension of each head
+matrix = lambda nout, nin, std=0.08: [[Value(random.gauss(0, std)) for _ in range(nin)] for _ in range(nout)]
+state_dict = {'wte': matrix(vocab_size, n_embd), 'wpe': matrix(block_size, n_embd), 'lm_head': matrix(vocab_size, n_embd)}
+for i in range(n_layer):
+    state_dict[f'layer{i}.attn_wq'] = matrix(n_embd, n_embd)
+    state_dict[f'layer{i}.attn_wk'] = matrix(n_embd, n_embd)
+    state_dict[f'layer{i}.attn_wv'] = matrix(n_embd, n_embd)
+    state_dict[f'layer{i}.attn_wo'] = matrix(n_embd, n_embd)
+    state_dict[f'layer{i}.mlp_fc1'] = matrix(4 * n_embd, n_embd)
+    state_dict[f'layer{i}.mlp_fc2'] = matrix(n_embd, 4 * n_embd)
+params = [p for mat in state_dict.values() for row in mat for p in row] # flatten params into a single list[Value]
+print(f"num params: {len(params)}")
+# Define the model architecture: a stateless function mapping token sequence and parameters to logits over what comes next.
+# Follow GPT-2, blessed among the GPTs, with minor differences: layernorm -> rmsnorm, no biases, GeLU -> ReLU
+
+def linear(x, w):
+    return [sum(wi * xi for wi, xi in zip(wo, x)) for wo in w]
+
+
+def softmax(logits):
+    max_val = max(val.data for val in logits)
+    exps = [(val - max_val).exp() for val in logits]
+    total = sum(exps)
+    return [e / total for e in exps]
+
+def rmsnorm(x):
+    ms = sum(xi * xi for xi in x) / len(x)
+    scale = (ms + 1e-5) ** -0.5
+    return [xi * scale for xi in x]
+
+def gpt(token_id, pos_id, keys, values):
+    tok_emb = state_dict['wte'][token_id] # token embedding
+    pos_emb = state_dict['wpe'][pos_id] # position embedding
+    x = [t + p for t, p in zip(tok_emb, pos_emb)] # joint token and position embedding
+    x = rmsnorm(x)
+
+    for li in range(n_layer):
+        # 1) Multi-head attention block
+        x_residual = x
+        x = rmsnorm(x)
+        q = linear(x, state_dict[f'layer{li}.attn_wq'])
+        k = linear(x, state_dict[f'layer{li}.attn_wk'])
+        v = linear(x, state_dict[f'layer{li}.attn_wv'])
+        keys[li].append(k)
+        values[li].append(v)
+        x_attn = []
+        for h in range(n_head):
+            hs = h * head_dim
+            q_h = q[hs:hs+head_dim]
+            k_h = [ki[hs:hs+head_dim] for ki in keys[li]]
+            v_h = [vi[hs:hs+head_dim] for vi in values[li]]
+            attn_logits = [sum(q_h[j] * k_h[t][j] for j in range(head_dim)) / head_dim**0.5 for t in range(len(k_h))]
+            attn_weights = softmax(attn_logits)
+            head_out = [sum(attn_weights[t] * v_h[t][j] for t in range(len(v_h))) for j in range(head_dim)]
+            x_attn.extend(head_out)
+        x = linear(x_attn, state_dict[f'layer{li}.attn_wo'])
+        x = [a + b for a, b in zip(x, x_residual)]
+        # 2) MLP block
+        x_residual = x
+        x = rmsnorm(x)
+        x = linear(x, state_dict[f'layer{li}.mlp_fc1'])
+        x = [xi.relu() for xi in x]
+        x = linear(x, state_dict[f'layer{li}.mlp_fc2'])
+        x = [a + b for a, b in zip(x, x_residual)]
+
+    logits = linear(x, state_dict['lm_head'])
+    return logits
+
+# Let there be Adam, the blessed optimizer and its buffers
+learning_rate, beta1, beta2, eps_adam = 0.01, 0.85, 0.99, 1e-8
+m = [0.0] * len(params) # first moment buffer
+v = [0.0] * len(params) # second moment buffer
+# Repeat in sequence
+num_steps = 1000 # number of training steps
+for step in range(num_steps):
+
+    # Take single document, tokenize it, surround it with BOS special token on both sides
+    doc = docs[step % len(docs)]
+    tokens = [BOS] + [uchars.index(ch) for ch in doc] + [BOS]
+    n = min(block_size, len(tokens) - 1)
+
+    # Forward the token sequence through the model, building up the computation graph all the way to the loss.
+    keys, values = [[] for _ in range(n_layer)], [[] for _ in range(n_layer)]
+    losses = []
+    for pos_id in range(n):
+        token_id, target_id = tokens[pos_id], tokens[pos_id + 1]
+        logits = gpt(token_id, pos_id, keys, values)
+        probs = softmax(logits)
+        loss_t = -probs[target_id].log()
+        losses.append(loss_t)
+    loss = (1 / n) * sum(losses) # final average loss over the document sequence. May yours be low.
+
+    # Backward the loss, calculating the gradients with respect to all model parameters.
+    loss.backward()
+
+    # Adam optimizer update: update the model parameters based on the corresponding gradients.
+    lr_t = learning_rate * (1 - step / num_steps) # linear learning rate decay
+    for i, p in enumerate(params):
+        m[i] = beta1 * m[i] + (1 - beta1) * p.grad
+        v[i] = beta2 * v[i] + (1 - beta2) * p.grad ** 2
+        m_hat = m[i] / (1 - beta1 ** (step + 1))
+        v_hat = v[i] / (1 - beta2 ** (step + 1))
+        p.data -= lr_t * m_hat / (v_hat ** 0.5 + eps_adam)
+        p.grad = 0
+
+    print(f"step {step+1:4d} / {num_steps:4d} | loss {loss.data:.4f}")
+
+# Inference: may the model babble back to us
+temperature = 0.5 # in (0, 1], control the "creativity" of generated text, low to high
+print("\n--- inference (new, hallucinated names) ---")
+for sample_idx in range(20):
+    keys, values = [[] for _ in range(n_layer)], [[] for _ in range(n_layer)]
+    token_id = BOS
+    sample = []
+    for pos_id in range(block_size):
+        logits = gpt(token_id, pos_id, keys, values)
+        probs = softmax([l / temperature for l in logits])
+        token_id = random.choices(range(vocab_size), weights=[p.data for p in probs])[0]
+        if token_id == BOS:
+            break
+        sample.append(uchars[token_id])
+    print(f"sample {sample_idx+1:2d}: {''.join(sample)}")