Add Voxtral MLX streaming backend

Integrates the voxmlx-based Voxtral Mini Realtime streaming pipeline:
- VoxtralStreamingASR and VoxtralStreamingOnlineProcessor
- Incremental audio encoding and token-by-token autoregressive decoding
- Selectable via --backend voxtral-mlx

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Dockerfile

@@ -37,9 +37,10 @@ RUN pip3 install --upgrade pip setuptools wheel && \
 COPY . .
 
 # Install WhisperLiveKit directly, allowing for optional dependencies
+# Example: --build-arg EXTRAS="translation"
 RUN if [ -n "$EXTRAS" ]; then \
       echo "Installing with extras: [$EXTRAS]"; \
-      pip install --no-cache-dir whisperlivekit[$EXTRAS]; \
+      pip install --no-cache-dir "whisperlivekit[$EXTRAS]"; \
     else \
       echo "Installing base package only"; \
       pip install --no-cache-dir whisperlivekit; \

README.md

@@ -1,24 +1,26 @@
-<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:
 
-- 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)
+- 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
@@ -51,9 +53,11 @@ 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.
 
@@ -96,11 +100,13 @@ wlk --host 0.0.0.0 --port 80 --model medium --diarization --language fr
 **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 +145,15 @@ 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/simul_whisper/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` |
+| `--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 +161,7 @@ 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 +171,7 @@ 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 +189,7 @@ 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. |
 
 
 
@@ -261,7 +267,7 @@ 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!
+  - `EXTRAS="translation"` - 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
 

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...

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.
 
 

docs/alignement_principles.md

@@ -0,0 +1,71 @@
+### Alignment between STT Tokens and Diarization Segments 
+
+- Example 1: The punctuation from STT and the speaker change from Diariation come in the prediction `t`
+- Example 2: The punctuation from STT comes from prediction `t`, but the speaker change from Diariation come in the prediction `t-1`
+- Example 3: The punctuation from STT comes from prediction `t-1`, but the speaker change from Diariation come in the prediction `t`
+
+> `#` Is the split between the `t-1` prediction and `t` prediction.  
+
+
+## Example 1:
+```text
+punctuations_segments : __#_______.__________________!____
+diarization_segments:
+SPK1                    __#____________
+SPK2                      #            ___________________
+-->
+ALIGNED SPK1            __#_______.
+ALIGNED SPK2              #        __________________!____
+
+t-1 output:
+SPK1:                   __#
+SPK2: NO
+DIARIZATION BUFFER: NO
+
+t output:
+SPK1:                       __#__.
+SPK2:                             __________________!____
+DIARIZATION BUFFER: No
+```
+
+## Example 2:
+```text
+punctuations_segments : _____#__.___________
+diarization_segments:
+SPK1                    ___  #
+SPK2                       __#______________
+-->
+ALIGNED SPK1            _____#__.
+ALIGNED SPK2                 #   ___________
+
+t-1 output:
+SPK1:                   ___  #
+SPK2:
+DIARIZATION BUFFER:        __#
+
+t output:
+SPK1:                      __#__.
+SPK2:                            ___________
+DIARIZATION BUFFER: No
+```
+
+## Example 3:
+```text
+punctuations_segments : ___.__#__________
+diarization_segments:
+SPK1                    ______#__
+SPK2                          #  ________
+-->
+ALIGNED SPK1            ___.  #
+ALIGNED SPK2                __#__________
+
+t-1 output:
+SPK1:                   ___.  #
+SPK2:
+DIARIZATION BUFFER:         __#
+
+t output:
+SPK1:                         #
+SPK2:                       __#___________
+DIARIZATION BUFFER: NO
+```

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

@@ -0,0 +1,43 @@
+# Technical Integration Guide
+
+This document introduce how to reuse the core components when you do **not** want to ship the bundled frontend, FastAPI server, or even the provided CLI.
+
+---
+
+## 1. Runtime Components
+
+| Layer | File(s) | Purpose |
+|-------|---------|---------|
+| Transport | `whisperlivekit/basic_server.py`, any ASGI/WebSocket server | Accepts audio over WebSocket (MediaRecorder WebM or raw PCM chunks) and streams JSON updates back |
+| Audio processing | `whisperlivekit/audio_processor.py` | Buffers audio, orchestrates transcription, diarization, translation, handles FFmpeg/PCM input |
+| Engines | `whisperlivekit/core.py`, `whisperlivekit/simul_whisper/*`, `whisperlivekit/local_agreement/*` | Load models once (SimulStreaming or LocalAgreement), expose `TranscriptionEngine` and helpers |
+| Frontends | `whisperlivekit/web/*`, `chrome-extension/*` | Optional UI layers feeding the WebSocket endpoint |
+
+**Key idea:** The server boundary is just `AudioProcessor.process_audio()` for incoming bytes and the async generator returned by `AudioProcessor.create_tasks()` for outgoing updates (`FrontData`). Everything else is optional.
+
+---
+
+## 2. Running Without the Bundled Frontend
+
+1. Start the server/engine however you like:
+   ```bash
+   wlk --model small --language en --host 0.0.0.0 --port 9000
+   # or launch your own app that instantiates TranscriptionEngine(...)
+   ```
+2. Build your own client (browser, mobile, desktop) that:
+   - Opens `ws(s)://<host>:<port>/asr`
+   - Sends either MediaRecorder/Opus WebM blobs **or** raw PCM (`--pcm-input` on the server tells the client to use the AudioWorklet).
+   - Consumes the JSON payload defined in `docs/API.md`.
+
+---
+
+## 3. Running Without FastAPI
+
+`whisperlivekit/basic_server.py` is just an example. Any async framework works, as long as you:
+
+1. Create a global `TranscriptionEngine` (expensive to initialize; reuse it).
+2. Instantiate `AudioProcessor(transcription_engine=engine)` for each connection.
+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.

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

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "whisperlivekit"
-version = "0.2.14"
+version = "0.2.18"
 description = "Real-time speech-to-text with speaker diarization using Whisper"
 readme = "README.md"
 authors = [
@@ -35,6 +35,7 @@ dependencies = [
     "torchaudio>=2.0.0",
     "torch>=2.0.0",
     "huggingface-hub>=0.25.0",
+    "faster-whisper>=1.2.0",
     "tqdm",
     "tiktoken",
     'triton>=2.0.0; platform_machine == "x86_64" and (sys_platform == "linux" or sys_platform == "linux2")'
@@ -56,15 +57,17 @@ 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.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"]

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/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")

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

@@ -1,58 +1,67 @@
 import asyncio
-import numpy as np
-from time import time, sleep
-import math
 import logging
 import traceback
-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 whisperlivekit.results_formater import format_output
+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.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)
 
 SENTINEL = object() # unique sentinel object for end of stream marker
+MIN_DURATION_REAL_SILENCE = 5
 
-def cut_at(cumulative_pcm, cut_sec):
-    cumulative_len = 0
-    cut_sample = int(cut_sec * 16000)
-    
-    for ind, pcm_array in enumerate(cumulative_pcm):
-        if (cumulative_len + len(pcm_array)) >= cut_sample:
-            cut_chunk = cut_sample - cumulative_len
-            before = np.concatenate(cumulative_pcm[:ind] + [cumulative_pcm[ind][:cut_chunk]])
-            after = [cumulative_pcm[ind][cut_chunk:]] + cumulative_pcm[ind+1:]
-            return before, after
-        cumulative_len += len(pcm_array)
-    return np.concatenate(cumulative_pcm), []
+async def get_all_from_queue(queue: asyncio.Queue) -> Union[object, Silence, np.ndarray, List[Any]]:
+    items: List[Any] = []
 
-async def get_all_from_queue(queue):
-    items = []
-    try:
-        while True:
-            item = queue.get_nowait()
-            items.append(item)
-    except asyncio.QueueEmpty:
-        pass
-    return items
+    first_item = await queue.get()
+    queue.task_done()
+    if first_item is SENTINEL:
+        return first_item
+    if isinstance(first_item, Silence):
+        return first_item
+    items.append(first_item)
+
+    while True:
+        if not queue._queue:
+            break
+        next_item = queue._queue[0]
+        if next_item is SENTINEL:
+            break
+        if isinstance(next_item, Silence):
+            break
+        items.append(await queue.get())
+        queue.task_done()
+    if isinstance(items[0], np.ndarray):
+        return np.concatenate(items)
+    else: #translation
+        return items
 
 class AudioProcessor:
     """
     Processes audio streams for transcription and diarization.
     Handles audio processing, state management, and result formatting.
     """
-    
-    def __init__(self, **kwargs):
+
+    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
@@ -64,36 +73,29 @@ class AudioProcessor:
         self.is_pcm_input = self.args.pcm_input
 
         # State management
-        self.is_stopping = False
-        self.silence = False
-        self.silence_duration = 0.0
-        self.state = State()
-        self.lock = asyncio.Lock()
-        self.sep = " "  # Default separator
-        self.last_response_content = FrontData()
-        self.last_detected_speaker = None
-        self.speaker_languages = {}
-        self.diarization_before_transcription = False
+        self.is_stopping: bool = False
+        self.current_silence: Optional[Silence] = None
+        self.state: State = State()
+        self.lock: asyncio.Lock = asyncio.Lock()
+        self.sep: str = " "  # Default separator
+        self.last_response_content: FrontData = FrontData()
 
-        self.segments = []
-        
+        self.tokens_alignment: TokensAlignment = TokensAlignment(self.state, self.args, self.sep)
+        self.beg_loop: Optional[float] = None
 
-        if self.diarization_before_transcription:
-            self.cumulative_pcm = []
-            self.last_start = 0.0
-            self.last_end = 0.0
-        
         # Models and processing
-        self.asr = models.asr
-        self.vac_model = models.vac_model
+        self.asr: Any = models.asr
+        self.vac: Optional[FixedVADIterator] = None
+
         if self.args.vac:
-            self.vac = FixedVADIterator(models.vac_model)
-        else:
-            self.vac = None
-                         
-        self.ffmpeg_manager = None
-        self.ffmpeg_reader_task = None
-        self._ffmpeg_error = 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
 
         if not self.is_pcm_input:
             self.ffmpeg_manager = FFmpegManager(
@@ -104,63 +106,103 @@ 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 = asyncio.Queue() if self.args.transcription else None
-        self.diarization_queue = asyncio.Queue() if self.args.diarization else None
-        self.translation_queue = asyncio.Queue() if self.args.target_language else None
-        self.pcm_buffer = bytearray()
 
-        self.transcription_task = None
-        self.diarization_task = None
-        self.translation_task = None
-        self.watchdog_task = None
-        self.all_tasks_for_cleanup = []
-        
-        self.transcription = None
-        self.translation = None
-        self.diarization = None
+        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
+        self.pcm_buffer: bytearray = bytearray()
+        self.total_pcm_samples: int = 0
+        self.transcription_task: Optional[asyncio.Task] = None
+        self.diarization_task: Optional[asyncio.Task] = None
+        self.translation_task: Optional[asyncio.Task] = None
+        self.watchdog_task: Optional[asyncio.Task] = None
+        self.all_tasks_for_cleanup: List[asyncio.Task] = []
+
+        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:
             self.translation = online_translation_factory(self.args, models.translation_model)
 
-    def convert_pcm_to_float(self, pcm_buffer):
+    async def _push_silence_event(self) -> None:
+        if self.transcription_queue:
+            await self.transcription_queue.put(self.current_silence)
+        if self.args.diarization and self.diarization_queue:
+            await self.diarization_queue.put(self.current_silence)
+        if self.translation_queue:
+            await self.translation_queue.put(self.current_silence)
+
+    async def _begin_silence(self) -> None:
+        if self.current_silence:
+            return
+        now = time() - self.beg_loop
+        self.current_silence = Silence(
+            is_starting=True, start=now
+        )
+        await self._push_silence_event()
+
+    async def _end_silence(self) -> 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
+        self.current_silence.compute_duration()
+        if self.current_silence.duration > MIN_DURATION_REAL_SILENCE:
+            self.state.new_tokens.append(self.current_silence)
+        await self._push_silence_event()
+        self.current_silence = None
+
+    async def _enqueue_active_audio(self, pcm_chunk: np.ndarray) -> None:
+        if pcm_chunk is None or pcm_chunk.size == 0:
+            return
+        if self.transcription_queue:
+            await self.transcription_queue.put(pcm_chunk.copy())
+        if self.args.diarization and self.diarization_queue:
+            await self.diarization_queue.put(pcm_chunk.copy())
+
+    def _slice_before_silence(self, pcm_array: np.ndarray, chunk_sample_start: int, silence_sample: Optional[int]) -> Optional[np.ndarray]:
+        if silence_sample is None:
+            return None
+        relative_index = int(silence_sample - chunk_sample_start)
+        if relative_index <= 0:
+            return None
+        split_index = min(relative_index, len(pcm_array))
+        if split_index <= 0:
+            return None
+        return pcm_array[:split_index]
+
+    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 add_dummy_token(self):
-        """Placeholder token when no transcription is available."""
-        async with self.lock:
-            current_time = time() - self.state.beg_loop
-            self.state.tokens.append(ASRToken(
-                start=current_time, end=current_time + 1,
-                text=".", speaker=-1, is_dummy=True
-            ))
-            
-    async def get_current_state(self):
+    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.state.beg_loop - self.state.end_buffer, 1))
-                
+                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):
+    async def ffmpeg_stdout_reader(self) -> None:
         """Read audio data from FFmpeg stdout and process it into the PCM pipeline."""
         beg = time()
         while True:
@@ -203,50 +245,60 @@ class AudioProcessor:
                 await asyncio.sleep(0.2)
 
         logger.info("FFmpeg stdout processing finished. Signaling downstream processors if needed.")
-        if not self.diarization_before_transcription and self.transcription_queue:
+        if self.transcription_queue:
             await self.transcription_queue.put(SENTINEL)
         if self.diarization:
             await self.diarization_queue.put(SENTINEL)
         if self.translation:
             await self.translation_queue.put(SENTINEL)
 
-    async def transcription_processor(self):
+    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 self.transcription_queue.get()
+                item = await get_all_from_queue(self.transcription_queue)
                 if item is SENTINEL:
                     logger.debug("Transcription processor received sentinel. Finishing.")
-                    self.transcription_queue.task_done()
                     break
 
                 asr_internal_buffer_duration_s = len(getattr(self.transcription, 'audio_buffer', [])) / self.transcription.SAMPLING_RATE
-                transcription_lag_s = max(0.0, time() - self.state.beg_loop - self.state.end_buffer)
+                transcription_lag_s = max(0.0, time() - self.beg_loop - self.state.end_buffer)
                 asr_processing_logs = f"internal_buffer={asr_internal_buffer_duration_s:.2f}s | lag={transcription_lag_s:.2f}s |"
-                if type(item) is Silence:
-                    asr_processing_logs += f" + Silence of = {item.duration:.2f}s"
+                stream_time_end_of_current_pcm = cumulative_pcm_duration_stream_time
+                new_tokens = []
+                current_audio_processed_upto = self.state.end_buffer
+
+                if isinstance(item, Silence):
+                    if item.is_starting:
+                        new_tokens, current_audio_processed_upto = await asyncio.to_thread(
+                            self.transcription.start_silence
+                        )
+                        asr_processing_logs += f" + Silence starting"
+                    if item.has_ended:
+                        asr_processing_logs += f" + Silence of = {item.duration:.2f}s"
+                        cumulative_pcm_duration_stream_time += item.duration
+                        current_audio_processed_upto = cumulative_pcm_duration_stream_time
+                        self.transcription.end_silence(item.duration, self.state.tokens[-1].end if self.state.tokens else 0)
                     if self.state.tokens:
                         asr_processing_logs += f" | last_end = {self.state.tokens[-1].end} |"
                     logger.info(asr_processing_logs)
-                    cumulative_pcm_duration_stream_time += item.duration
-                    self.transcription.insert_silence(item.duration, self.state.tokens[-1].end if self.state.tokens else 0)
-                    continue
+                    new_tokens = new_tokens or []
+                    current_audio_processed_upto = max(current_audio_processed_upto, stream_time_end_of_current_pcm)
                 elif isinstance(item, ChangeSpeaker):
                     self.transcription.new_speaker(item)
+                    continue
                 elif isinstance(item, np.ndarray):
                     pcm_array = item
-                
-                logger.info(asr_processing_logs)
-                
-                duration_this_chunk = len(pcm_array) / self.sample_rate
-                cumulative_pcm_duration_stream_time += duration_this_chunk
-                stream_time_end_of_current_pcm = cumulative_pcm_duration_stream_time
+                    logger.info(asr_processing_logs)
+                    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)
+                    new_tokens, current_audio_processed_upto = await asyncio.to_thread(self.transcription.process_iter)
+                    new_tokens = new_tokens or []
 
-                self.transcription.insert_audio_chunk(pcm_array, stream_time_end_of_current_pcm)
-                new_tokens, current_audio_processed_upto = await asyncio.to_thread(self.transcription.process_iter)
-                
                 _buffer_transcript = self.transcription.get_buffer()
                 buffer_text = _buffer_transcript.text
 
@@ -259,29 +311,28 @@ 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
                     self.state.end_buffer = max(candidate_end_times)
-                
+                    self.state.new_tokens.extend(new_tokens)
+                    self.state.new_tokens_buffer = _buffer_transcript
+
                 if self.translation_queue:
                     for token in new_tokens:
                         await self.translation_queue.put(token)
-                        
-                self.transcription_queue.task_done()
-                
             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:
@@ -292,124 +343,60 @@ class AudioProcessor:
         logger.info("Transcription processor task finished.")
 
 
-    async def diarization_processor(self, diarization_obj):
-        """Process audio chunks for speaker diarization."""
-        if self.diarization_before_transcription:
-            self.current_speaker = 0
-            await self.transcription_queue.put(ChangeSpeaker(speaker=self.current_speaker, start=0.0))
+    async def diarization_processor(self) -> None:
         while True:
             try:
-                item = await self.diarization_queue.get()
+                item = await get_all_from_queue(self.diarization_queue)
                 if item is SENTINEL:
-                    logger.debug("Diarization processor received sentinel. Finishing.")
-                    self.diarization_queue.task_done()
                     break
                 elif type(item) is Silence:
-                    diarization_obj.insert_silence(item.duration)
+                    if item.has_ended:
+                        self.diarization.insert_silence(item.duration)
                     continue
-                elif isinstance(item, np.ndarray):
-                    pcm_array = item
-                else:
-                    raise Exception('item should be pcm_array') 
-                
-                
-                
-                # Process diarization
-                await diarization_obj.diarize(pcm_array)
-                if self.diarization_before_transcription:
-                    segments = diarization_obj.get_segments()
-                    self.cumulative_pcm.append(pcm_array)
-                    if segments:
-                        last_segment = segments[-1]                    
-                        if last_segment.speaker != self.current_speaker:
-                            cut_sec = last_segment.start - self.last_end
-                            to_transcript, self.cumulative_pcm = cut_at(self.cumulative_pcm, cut_sec)
-                            await self.transcription_queue.put(to_transcript)
-                            
-                            self.current_speaker = last_segment.speaker
-                            await self.transcription_queue.put(ChangeSpeaker(speaker=self.current_speaker, start=last_segment.start))
-                            
-                            cut_sec = last_segment.end - last_segment.start
-                            to_transcript, self.cumulative_pcm = cut_at(self.cumulative_pcm, cut_sec)
-                            await self.transcription_queue.put(to_transcript)                            
-                            self.last_start = last_segment.start
-                            self.last_end = last_segment.end
-                        else:
-                            cut_sec = last_segment.end - self.last_end
-                            to_transcript, self.cumulative_pcm = cut_at(self.cumulative_pcm, cut_sec)
-                            await self.transcription_queue.put(to_transcript)
-                            self.last_end = last_segment.end
-                elif not self.diarization_before_transcription:           
-                    async with self.lock:
-                        self.state.tokens = diarization_obj.assign_speakers_to_tokens(
-                            self.state.tokens,
-                            use_punctuation_split=self.args.punctuation_split
-                        )
-                if len(self.state.tokens) > 0:
-                    self.state.end_attributed_speaker = max(self.state.tokens[-1].end, self.state.end_attributed_speaker)
-                self.diarization_queue.task_done()
-                
+                self.diarization.insert_audio_chunk(item)
+                diarization_segments = await self.diarization.diarize()
+                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()}")
-                if 'pcm_array' in locals() and pcm_array is not SENTINEL:
-                    self.diarization_queue.task_done()
         logger.info("Diarization processor task finished.")
 
-    async def translation_processor(self):
-        # the idea is to ignore diarization for the moment. We use only transcription tokens. 
+    async def translation_processor(self) -> None:
+        # 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:
             try:
-                item = await self.translation_queue.get() #block until at least 1 token
+                item = await get_all_from_queue(self.translation_queue)
                 if item is SENTINEL:
                     logger.debug("Translation processor received sentinel. Finishing.")
-                    self.translation_queue.task_done()
                     break
                 elif type(item) is Silence:
-                    self.translation.insert_silence(item.duration)
-                    continue
-                
-                # get all the available tokens for translation. The more words, the more precise
-                tokens_to_process = [item]
-                additional_tokens = await get_all_from_queue(self.translation_queue)
-                
-                sentinel_found = False
-                for additional_token in additional_tokens:
-                    if additional_token is SENTINEL:
-                        sentinel_found = True
-                        break
-                    elif type(additional_token) is Silence:
-                        self.translation.insert_silence(additional_token.duration)
+                    if item.is_starting:
+                        new_translation, new_translation_buffer = self.translation.validate_buffer_and_reset()
+                    if item.has_ended:
+                        self.translation.insert_silence(item.duration)
                         continue
-                    else:
-                        tokens_to_process.append(additional_token)                
-                if tokens_to_process:
-                    self.translation.insert_tokens(tokens_to_process)
-                    translation_validated_segments, buffer_translation = await asyncio.to_thread(self.translation.process)
-                    async with self.lock:
-                        self.state.translation_validated_segments = translation_validated_segments
-                        self.state.buffer_translation = buffer_translation
-                self.translation_queue.task_done()
-                for _ in additional_tokens:
-                    self.translation_queue.task_done()
-                
-                if sentinel_found:
-                    logger.debug("Translation processor received sentinel in batch. Finishing.")
-                    break
-                
+                elif isinstance(item, ChangeSpeaker):
+                    new_translation, new_translation_buffer = self.translation.validate_buffer_and_reset()
+                    pass
+                else:
+                    self.translation.insert_tokens(item)
+                    new_translation, new_translation_buffer = await asyncio.to_thread(self.translation.process)
+                async with self.lock:
+                    self.state.new_translation.append(new_translation)
+                    self.state.new_translation_buffer = new_translation_buffer
             except Exception as e:
                 logger.warning(f"Exception in translation_processor: {e}")
                 logger.warning(f"Traceback: {traceback.format_exc()}")
-                if 'token' in locals() and item is not SENTINEL:
-                    self.translation_queue.task_done()
-                if 'additional_tokens' in locals():
-                    for _ in additional_tokens:
-                        self.translation_queue.task_done()
         logger.info("Translation processor task finished.")
 
-    async def results_formatter(self):
+    async def results_formatter(self) -> AsyncGenerator[FrontData, None]:
         """Format processing results for output."""
         while True:
             try:
@@ -419,79 +406,56 @@ class AudioProcessor:
                     await asyncio.sleep(1)
                     continue
 
-                state = await self.get_current_state()
-                
-                lines, undiarized_text = format_output(
-                    state,
-                    self.silence,
-                    args = self.args,
-                    sep=self.sep
+                self.tokens_alignment.update()
+                lines, buffer_diarization_text, buffer_translation_text = self.tokens_alignment.get_lines(
+                    diarization=self.args.diarization,
+                    translation=bool(self.translation),
+                    current_silence=self.current_silence
                 )
-                if lines and lines[-1].speaker == -2:
-                    buffer_transcription = Transcript()
-                else:
-                    buffer_transcription = state.buffer_transcription
+                state = await self.get_current_state()
 
-                buffer_diarization = ''
-                if undiarized_text:
-                    buffer_diarization = self.sep.join(undiarized_text)
+                buffer_transcription_text = state.buffer_transcription.text if state.buffer_transcription else ''
 
-                    async with self.lock:
-                        self.state.end_attributed_speaker = state.end_attributed_speaker
-
-                buffer_translation_text = ''
-                if state.buffer_translation:
-                    raw_buffer_translation = getattr(state.buffer_translation, 'text', state.buffer_translation)
-                    if raw_buffer_translation:
-                        buffer_translation_text = raw_buffer_translation.strip()
-                
                 response_status = "active_transcription"
-                if not state.tokens and not buffer_transcription and not buffer_diarization:
+                if not lines and not buffer_transcription_text and not buffer_diarization_text:
                     response_status = "no_audio_detected"
-                    lines = []
-                elif not lines:
-                    lines = [Line(
-                        speaker=1,
-                        start=state.end_buffer,
-                        end=state.end_buffer
-                    )]
-                
+
                 response = FrontData(
                     status=response_status,
                     lines=lines,
-                    buffer_transcription=buffer_transcription.text.strip(),
-                    buffer_diarization=buffer_diarization,
+                    buffer_transcription=buffer_transcription_text,
+                    buffer_diarization=buffer_diarization_text,
                     buffer_translation=buffer_translation_text,
                     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 and (lines or buffer_transcription or buffer_diarization or response_status == "no_audio_detected"):
+                if should_push:
                     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):
+
+    async def create_tasks(self) -> AsyncGenerator[FrontData, None]:
         """Create and start processing tasks."""
         self.all_tasks_for_cleanup = []
-        processing_tasks_for_watchdog = []
+        processing_tasks_for_watchdog: List[asyncio.Task] = []
 
         # If using FFmpeg (non-PCM input), start it and spawn stdout reader
         if not self.is_pcm_input:
             success = await self.ffmpeg_manager.start()
             if not success:
                 logger.error("Failed to start FFmpeg manager")
-                async def error_generator():
+                async def error_generator() -> AsyncGenerator[FrontData, None]:
                     yield FrontData(
                         status="error",
                         error="FFmpeg failed to start. Please check that FFmpeg is installed."
@@ -505,26 +469,26 @@ 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.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):
+    async def watchdog(self, tasks_to_monitor: List[asyncio.Task]) -> None:
         """Monitors the health of critical processing tasks."""
-        tasks_remaining = [task for task in tasks_to_monitor if task]
+        tasks_remaining: List[asyncio.Task] = [task for task in tasks_to_monitor if task]
         while True:
             try:
                 if not tasks_remaining:
@@ -532,7 +496,7 @@ class AudioProcessor:
                     return
 
                 await asyncio.sleep(10)
-                
+
                 for i, task in enumerate(list(tasks_remaining)):
                     if task.done():
                         exc = task.exception()
@@ -542,21 +506,21 @@ 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):
+
+    async def cleanup(self) -> None:
         """Clean up resources when processing is complete."""
         logger.info("Starting cleanup of AudioProcessor resources.")
         self.is_stopping = True
         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)
@@ -572,7 +536,7 @@ class AudioProcessor:
             self.diarization.close()
         logger.info("AudioProcessor cleanup complete.")
 
-    def _processing_tasks_done(self):
+    def _processing_tasks_done(self) -> bool:
         """Return True when all active processing tasks have completed."""
         tasks_to_check = [
             self.transcription_task,
@@ -583,16 +547,18 @@ class AudioProcessor:
         return all(task.done() for task in tasks_to_check if task)
 
 
-    async def process_audio(self, message):
+    async def process_audio(self, message: Optional[bytes]) -> None:
         """Process incoming audio data."""
 
-        if not self.state.beg_loop:
-            self.state.beg_loop = time()
+        if not self.beg_loop:
+            self.beg_loop = time()
+            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
-             
+
             if self.transcription_queue:
                 await self.transcription_queue.put(SENTINEL)
 
@@ -620,7 +586,7 @@ class AudioProcessor:
                 else:
                     logger.warning("Failed to write audio data to FFmpeg")
 
-    async def handle_pcm_data(self):
+    async def handle_pcm_data(self) -> None:
         # Process when enough data
         if len(self.pcm_buffer) < self.bytes_per_sec:
             return
@@ -633,46 +599,36 @@ 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])
         self.pcm_buffer = self.pcm_buffer[aligned_chunk_size:]
 
-        res = None
-        end_of_audio = False
-        silence_buffer = None
+        num_samples = len(pcm_array)
+        chunk_sample_start = self.total_pcm_samples
+        chunk_sample_end = chunk_sample_start + num_samples
 
+        res = None
         if self.args.vac:
             res = self.vac(pcm_array)
 
         if res is not None:
-            if res.get("end", 0) > res.get("start", 0):
-                end_of_audio = True
-            elif self.silence: #end of silence
-                self.silence = False
-                silence_buffer = Silence(duration=time() - self.start_silence)
+            if "start" in res and self.current_silence:
+                await self._end_silence()
 
-        if silence_buffer:
-            if not self.diarization_before_transcription and self.transcription_queue:
-                await self.transcription_queue.put(silence_buffer)
-            if self.args.diarization and self.diarization_queue:
-                await self.diarization_queue.put(silence_buffer)
-            if self.translation_queue:
-                await self.translation_queue.put(silence_buffer)
+            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()
 
-        if not self.silence:
-            if not self.diarization_before_transcription and self.transcription_queue:
-                await self.transcription_queue.put(pcm_array.copy())
+        if not self.current_silence:
+            await self._enqueue_active_audio(pcm_array)
 
-            if self.args.diarization and self.diarization_queue:
-                await self.diarization_queue.put(pcm_array.copy())
-
-            self.silence_duration = 0.0
-
-            if end_of_audio:
-                self.silence = True
-                self.start_silence = time()
+        self.total_pcm_samples = chunk_sample_end
 
         if not self.args.transcription and not self.args.diarization:
             await asyncio.sleep(0.1)

whisperlivekit/backend_support.py

@@ -29,6 +29,13 @@ def mlx_backend_available(warn_on_missing = False):
     return available
 
 
+def voxmlx_backend_available():
+    """Return True if voxmlx (Voxtral MLX backend) is available."""
+    is_macos = platform.system() == "Darwin"
+    is_arm = platform.machine() == "arm64"
+    return is_macos and is_arm and module_available("voxmlx")
+
+
 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

@@ -1,10 +1,13 @@
-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)

whisperlivekit/core.py

@@ -1,9 +1,12 @@
+import logging
+import sys
+import threading
+from argparse import Namespace
+
+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({
@@ -17,16 +20,26 @@ 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
+        # Thread-safe initialization check
+        with TranscriptionEngine._lock:
+            if TranscriptionEngine._initialized:
+                return
+            # Set flag immediately to prevent re-initialization
+            TranscriptionEngine._initialized = True
 
+        # Perform initialization outside lock to avoid holding lock during slow operations
         global_params = {
             "host": "localhost",
             "port": 8000,
@@ -34,7 +47,6 @@ class TranscriptionEngine:
             "punctuation_split": False,
             "target_language": "",
             "vac": True,
-            "vac_onnx": False,
             "vac_chunk_size": 0.04,
             "log_level": "DEBUG",
             "ssl_certfile": None,
@@ -52,11 +64,12 @@ class TranscriptionEngine:
 
         transcription_common_params = {
             "warmup_file": None,
-            "min_chunk_size": 0.5,
-            "model_size": "tiny",
+            "min_chunk_size": 0.1,
+            "model_size": "base",
             "model_cache_dir": None,
             "model_dir": None,
             "model_path": None,
+            "lora_path": None,
             "lan": "auto",
             "direct_english_translation": False,
         }
@@ -76,17 +89,27 @@ class TranscriptionEngine:
         self.asr = None
         self.tokenizer = None
         self.diarization = None
-        self.vac_model = None
+        self.vac_session = 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)
-        
+            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"
+                )
         backend_policy = self.args.backend_policy
         if self.args.transcription:
-            if backend_policy == "simulstreaming":                 
+            if self.args.backend == "voxtral-mlx":
+                from whisperlivekit.voxtral_streaming import VoxtralStreamingASR
+                self.tokenizer = None
+                self.asr = VoxtralStreamingASR(**transcription_common_params)
+                logger.info("Using Voxtral MLX streaming backend")
+            elif backend_policy == "simulstreaming":
                 simulstreaming_params = {
                     "disable_fast_encoder": False,
                     "custom_alignment_heads": None,
@@ -100,7 +123,6 @@ class TranscriptionEngine:
                     "init_prompt": None,
                     "static_init_prompt": None,
                     "max_context_tokens": None,
-                    "preload_model_count": 1,
                 }
                 simulstreaming_params = update_with_kwargs(simulstreaming_params, kwargs)
                 
@@ -135,7 +157,8 @@ class TranscriptionEngine:
 
         if self.args.diarization:
             if self.args.diarization_backend == "diart":
-                from whisperlivekit.diarization.diart_backend import DiartDiarization
+                from whisperlivekit.diarization.diart_backend import \
+                    DiartDiarization
                 diart_params = {
                     "segmentation_model": "pyannote/segmentation-3.0",
                     "embedding_model": "pyannote/embedding",
@@ -146,7 +169,8 @@ class TranscriptionEngine:
                     **diart_params
                 )
             elif self.args.diarization_backend == "sortformer":
-                from whisperlivekit.diarization.sortformer_backend import SortformerDiarization
+                from whisperlivekit.diarization.sortformer_backend import \
+                    SortformerDiarization
                 self.diarization_model = SortformerDiarization()
         
         self.translation_model = None
@@ -164,16 +188,16 @@ class TranscriptionEngine:
                 }
                 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
 
 
 def online_factory(args, asr):
-    if args.backend_policy == "simulstreaming":    
+    if getattr(args, 'backend', None) == "voxtral-mlx":
+        from whisperlivekit.voxtral_streaming import VoxtralStreamingOnlineProcessor
+        return VoxtralStreamingOnlineProcessor(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):
@@ -182,7 +206,8 @@ def online_diarization_factory(args, 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
+        from whisperlivekit.diarization.sortformer_backend import \
+            SortformerDiarizationOnline
         online = SortformerDiarizationOnline(shared_model=diarization_backend)
     return online
 

whisperlivekit/diarization/diart_backend.py

@@ -1,20 +1,20 @@
 import asyncio
+import logging
 import re
 import threading
-import numpy as np
-import logging
 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.timed_objects import SpeakerSegment
 
 logger = logging.getLogger(__name__)
 
@@ -26,7 +26,7 @@ class DiarizationObserver(Observer):
     """Observer that logs all data emitted by the diarization pipeline and stores speaker segments."""
     
     def __init__(self):
-        self.speaker_segments = []
+        self.diarization_segments = []
         self.processed_time = 0
         self.segment_lock = threading.Lock()
         self.global_time_offset = 0.0
@@ -48,7 +48,7 @@ class DiarizationObserver(Observer):
                 for speaker, label in annotation._labels.items():
                     for start, end in zip(label.segments_boundaries_[:-1], label.segments_boundaries_[1:]):
                         print(f"  {speaker}: {start:.2f}s-{end:.2f}s")
-                        self.speaker_segments.append(SpeakerSegment(
+                        self.diarization_segments.append(SpeakerSegment(
                             speaker=speaker,
                             start=start + self.global_time_offset,
                             end=end + self.global_time_offset
@@ -59,14 +59,14 @@ class DiarizationObserver(Observer):
     def get_segments(self) -> List[SpeakerSegment]:
         """Get a copy of the current speaker segments."""
         with self.segment_lock:
-            return self.speaker_segments.copy()
+            return self.diarization_segments.copy()
     
     def clear_old_segments(self, older_than: float = 30.0):
         """Clear segments older than the specified time."""
         with self.segment_lock:
             current_time = self.processed_time
-            self.speaker_segments = [
-                segment for segment in self.speaker_segments 
+            self.diarization_segments = [
+                segment for segment in self.diarization_segments 
                 if current_time - segment.end < older_than
             ]
     
@@ -178,7 +178,6 @@ class DiartDiarization:
         
         self.pipeline = SpeakerDiarization(config=config)        
         self.observer = DiarizationObserver()
-        self.lag_diart = None
         
         if use_microphone:
             self.source = MicrophoneAudioSource(block_duration=block_duration)
@@ -203,46 +202,20 @@ 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."""
         if self.custom_source:
             self.custom_source.close()
 
-    def assign_speakers_to_tokens(self, tokens: list, use_punctuation_split: bool = False) -> float:
-        """
-        Assign speakers to tokens based on timing overlap with speaker segments.
-        Uses the segments collected by the observer.
-        
-        If use_punctuation_split is True, uses punctuation marks to refine speaker boundaries.
-        """
-        segments = self.observer.get_segments()
-        
-        # Debug logging
-        logger.debug(f"assign_speakers_to_tokens called with {len(tokens)} tokens")
-        logger.debug(f"Available segments: {len(segments)}")
-        for i, seg in enumerate(segments[:5]):  # Show first 5 segments
-            logger.debug(f"  Segment {i}: {seg.speaker} [{seg.start:.2f}-{seg.end:.2f}]")
-        
-        if not self.lag_diart and segments and tokens:
-            self.lag_diart = segments[0].start - tokens[0].start
-        
-        if not use_punctuation_split:
-            for token in tokens:
-                for segment in segments:
-                    if not (segment.end <= token.start + self.lag_diart or segment.start >= token.end + self.lag_diart):
-                        token.speaker = extract_number(segment.speaker) + 1
-        else:
-            tokens = add_speaker_to_tokens(segments, tokens)
-        return tokens
         
 def concatenate_speakers(segments):
     segments_concatenated = [{"speaker": 1, "begin": 0.0, "end": 0.0}]

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
 
@@ -94,11 +95,11 @@ class SortformerDiarizationOnline:
             model_name: Pre-trained model name (default: "nvidia/diar_streaming_sortformer_4spk-v2")
         """
         self.sample_rate = sample_rate
-        self.speaker_segments = []
+        self.diarization_segments = []
+        self.diar_segments = []
         self.buffer_audio = np.array([], dtype=np.float32)
         self.segment_lock = threading.Lock()
         self.global_time_offset = 0.0
-        self.processed_time = 0.0
         self.debug = False
                 
         self.diar_model = shared_model.diar_model
@@ -155,12 +156,10 @@ class SortformerDiarizationOnline:
         )
         self.streaming_state.fifo_lengths = torch.zeros((batch_size,), dtype=torch.long, device=device)
         self.streaming_state.mean_sil_emb = torch.zeros((batch_size, self.diar_model.sortformer_modules.fc_d_model), device=device)
-        self.streaming_state.n_sil_frames = torch.zeros((batch_size,), dtype=torch.long, device=device)
-        
-        # Initialize total predictions tensor
+        self.streaming_state.n_sil_frames = torch.zeros((batch_size,), dtype=torch.long, device=device)        
         self.total_preds = torch.zeros((batch_size, 0, self.diar_model.sortformer_modules.n_spk), device=device)
 
-    def insert_silence(self, silence_duration: float):
+    def insert_silence(self, silence_duration: Optional[float]):
         """
         Insert silence period by adjusting the global time offset.
         
@@ -171,248 +170,111 @@ class SortformerDiarizationOnline:
             self.global_time_offset += silence_duration
         logger.debug(f"Inserted silence of {silence_duration:.2f}s, new offset: {self.global_time_offset:.2f}s")
 
-    async def diarize(self, pcm_array: np.ndarray):
+    def insert_audio_chunk(self, pcm_array: np.ndarray):
+        if self.debug:
+            self.audio_buffer.append(pcm_array.copy())
+        self.buffer_audio = np.concatenate([self.buffer_audio, pcm_array.copy()])
+  
+
+    async def diarize(self):
         """
         Process audio data for diarization in streaming fashion.
         
         Args:
             pcm_array: Audio data as numpy array
         """
-        try:
-            if self.debug:
-                self.audio_buffer.append(pcm_array.copy())
 
-            threshold = int(self.chunk_duration_seconds * self.sample_rate)
+        threshold = int(self.chunk_duration_seconds * self.sample_rate)
+        
+        if not len(self.buffer_audio) >= threshold:
+            return []
+        
+        audio = self.buffer_audio[:threshold]
+        self.buffer_audio = self.buffer_audio[threshold:]
+        
+        device = self.diar_model.device
+        audio_signal_chunk = torch.tensor(audio, device=device).unsqueeze(0)
+        audio_signal_length_chunk = torch.tensor([audio_signal_chunk.shape[1]], device=device)
+        
+        processed_signal_chunk, processed_signal_length_chunk = self.audio2mel.get_features(
+            audio_signal_chunk, audio_signal_length_chunk
+        )
+        processed_signal_chunk = processed_signal_chunk.to(device)
+        processed_signal_length_chunk = processed_signal_length_chunk.to(device)
+        
+        if self._previous_chunk_features is not None:
+            to_add = self._previous_chunk_features[:, :, -99:].to(device)
+            total_features = torch.concat([to_add, processed_signal_chunk], dim=2).to(device)
+        else:
+            total_features = processed_signal_chunk.to(device)
+        
+        self._previous_chunk_features = processed_signal_chunk.to(device)
+        
+        chunk_feat_seq_t = torch.transpose(total_features, 1, 2).to(device)
+        
+        with torch.inference_mode():
+            left_offset = 8 if self._chunk_index > 0 else 0
+            right_offset = 8
             
-            self.buffer_audio = np.concatenate([self.buffer_audio, pcm_array.copy()])
-            if not len(self.buffer_audio) >= threshold:
-                return
-            
-            audio = self.buffer_audio[:threshold]
-            self.buffer_audio = self.buffer_audio[threshold:]
-            
-            device = self.diar_model.device
-            audio_signal_chunk = torch.tensor(audio, device=device).unsqueeze(0)
-            audio_signal_length_chunk = torch.tensor([audio_signal_chunk.shape[1]], device=device)
-            
-            processed_signal_chunk, processed_signal_length_chunk = self.audio2mel.get_features(
-                audio_signal_chunk, audio_signal_length_chunk
-            )
-            processed_signal_chunk = processed_signal_chunk.to(device)
-            processed_signal_length_chunk = processed_signal_length_chunk.to(device)
-            
-            if self._previous_chunk_features is not None:
-                to_add = self._previous_chunk_features[:, :, -99:].to(device)
-                total_features = torch.concat([to_add, processed_signal_chunk], dim=2).to(device)
-            else:
-                total_features = processed_signal_chunk.to(device)
-            
-            self._previous_chunk_features = processed_signal_chunk.to(device)
-            
-            chunk_feat_seq_t = torch.transpose(total_features, 1, 2).to(device)
-            
-            with torch.inference_mode():
-                left_offset = 8 if self._chunk_index > 0 else 0
-                right_offset = 8
-                
-                self.streaming_state, self.total_preds = self.diar_model.forward_streaming_step(
-                    processed_signal=chunk_feat_seq_t,
-                    processed_signal_length=torch.tensor([chunk_feat_seq_t.shape[1]]).to(device),
-                    streaming_state=self.streaming_state,
-                    total_preds=self.total_preds,
-                    left_offset=left_offset,
-                    right_offset=right_offset,
-                )
-                
-            # Convert predictions to speaker segments
-            self._process_predictions()
-            
-            self._chunk_index += 1
-            
-        except Exception as e:
-            logger.error(f"Error in diarize: {e}")
-            raise
-            
-        # TODO: Handle case when stream ends with partial buffer (accumulated_duration > 0 but < chunk_duration_seconds)
+            self.streaming_state, self.total_preds = self.diar_model.forward_streaming_step(
+                processed_signal=chunk_feat_seq_t,
+                processed_signal_length=torch.tensor([chunk_feat_seq_t.shape[1]]).to(device),
+                streaming_state=self.streaming_state,
+                total_preds=self.total_preds,
+                left_offset=left_offset,
+                right_offset=right_offset,
+            )                
+        new_segments = self._process_predictions()
+        
+        self._chunk_index += 1
+        return new_segments
 
     def _process_predictions(self):
         """Process model predictions and convert to speaker segments."""
-        try:
-            preds_np = self.total_preds[0].cpu().numpy()
-            active_speakers = np.argmax(preds_np, axis=1)
-            
-            if self._len_prediction is None:
-                self._len_prediction = len(active_speakers)
-            
-            # Get predictions for current chunk
-            frame_duration = self.chunk_duration_seconds / self._len_prediction
-            current_chunk_preds = active_speakers[-self._len_prediction:]
-            
-            with self.segment_lock:
-                # Process predictions into segments
-                base_time = self._chunk_index * self.chunk_duration_seconds + self.global_time_offset
-                
-                for idx, spk in enumerate(current_chunk_preds):
-                    start_time = base_time + idx * frame_duration
-                    end_time = base_time + (idx + 1) * frame_duration
-                    
-                    # Check if this continues the last segment or starts a new one
-                    if (self.speaker_segments and 
-                        self.speaker_segments[-1].speaker == spk and 
-                        abs(self.speaker_segments[-1].end - start_time) < frame_duration * 0.5):
-                        # Continue existing segment
-                        self.speaker_segments[-1].end = end_time
-                    else:
-                        
-                        # Create new segment
-                        self.speaker_segments.append(SpeakerSegment(
-                            speaker=spk,
-                            start=start_time,
-                            end=end_time
-                        ))
-                
-                # Update processed time
-                self.processed_time = max(self.processed_time, base_time + self.chunk_duration_seconds)
-                
-                logger.debug(f"Processed chunk {self._chunk_index}, total segments: {len(self.speaker_segments)}")
-                
-        except Exception as e:
-            logger.error(f"Error processing predictions: {e}")
-
-    def assign_speakers_to_tokens(self, tokens: list, use_punctuation_split: bool = False) -> list:
-        """
-        Assign speakers to tokens based on timing overlap with speaker segments.
+        preds_np = self.total_preds[0].cpu().numpy()
+        active_speakers = np.argmax(preds_np, axis=1)
         
-        Args:
-            tokens: List of tokens with timing information
-            use_punctuation_split: Whether to use punctuation for boundary refinement
-            
-        Returns:
-            List of tokens with speaker assignments
-            Last speaker_segment
-        """
+        if self._len_prediction is None:
+            self._len_prediction = len(active_speakers) #12
+        
+        frame_duration = self.chunk_duration_seconds / self._len_prediction
+        current_chunk_preds = active_speakers[-self._len_prediction:]
+        
+        new_segments = []
+
         with self.segment_lock:
-            segments = self.speaker_segments.copy()
-        
-        if not segments or not tokens:
-            logger.debug("No segments or tokens available for speaker assignment")
-            return tokens
-        
-        logger.debug(f"Assigning speakers to {len(tokens)} tokens using {len(segments)} segments")
-        use_punctuation_split = False
-        if not use_punctuation_split:
-            # Simple overlap-based assignment
-            for token in tokens:
-                token.speaker = -1  # Default to no speaker
-                for segment in segments:
-                    # Check for timing overlap
-                    if not (segment.end <= token.start or segment.start >= token.end):
-                        token.speaker = segment.speaker + 1  # Convert to 1-based indexing
-                        break
-        else:
-            # Use punctuation-aware assignment (similar to diart_backend)
-            tokens = self._add_speaker_to_tokens_with_punctuation(segments, tokens)
-        
-        return tokens
-
-    def _add_speaker_to_tokens_with_punctuation(self, segments: List[SpeakerSegment], tokens: list) -> list:
-        """
-        Assign speakers to tokens with punctuation-aware boundary adjustment.
-        
-        Args:
-            segments: List of speaker segments
-            tokens: List of tokens to assign speakers to
-            
-        Returns:
-            List of tokens with speaker assignments
-        """
-        punctuation_marks = {'.', '!', '?'}
-        punctuation_tokens = [token for token in tokens if token.text.strip() in punctuation_marks]
-        
-        # Convert segments to concatenated format
-        segments_concatenated = self._concatenate_speakers(segments)
-        
-        # Adjust segment boundaries based on punctuation
-        for ind, segment in enumerate(segments_concatenated):
-            for i, punctuation_token in enumerate(punctuation_tokens):
-                if punctuation_token.start > segment['end']:
-                    after_length = punctuation_token.start - segment['end']
-                    before_length = segment['end'] - punctuation_tokens[i - 1].end if i > 0 else float('inf')
-                    
-                    if before_length > after_length:
-                        segment['end'] = punctuation_token.start
-                        if i < len(punctuation_tokens) - 1 and ind + 1 < len(segments_concatenated):
-                            segments_concatenated[ind + 1]['begin'] = punctuation_token.start
-                    else:
-                        segment['end'] = punctuation_tokens[i - 1].end if i > 0 else segment['end']
-                        if i < len(punctuation_tokens) - 1 and ind - 1 >= 0:
-                            segments_concatenated[ind - 1]['begin'] = punctuation_tokens[i - 1].end
-                    break
-        
-        # Ensure non-overlapping tokens
-        last_end = 0.0
-        for token in tokens:
-            start = max(last_end + 0.01, token.start)
-            token.start = start
-            token.end = max(start, token.end)
-            last_end = token.end
-        
-        # Assign speakers based on adjusted segments
-        ind_last_speaker = 0
-        for segment in segments_concatenated:
-            for i, token in enumerate(tokens[ind_last_speaker:]):
-                if token.end <= segment['end']:
-                    token.speaker = segment['speaker']
-                    ind_last_speaker = i + 1
-                elif token.start > segment['end']:
-                    break
-        
-        return tokens
-
-    def _concatenate_speakers(self, segments: List[SpeakerSegment]) -> List[dict]:
-        """
-        Concatenate consecutive segments from the same speaker.
-        
-        Args:
-            segments: List of speaker segments
-            
-        Returns:
-            List of concatenated speaker segments
-        """
-        if not segments:
-            return []
-            
-        segments_concatenated = [{"speaker": segments[0].speaker + 1, "begin": segments[0].start, "end": segments[0].end}]
-        
-        for segment in segments[1:]:
-            speaker = segment.speaker + 1
-            if segments_concatenated[-1]['speaker'] != speaker:
-                segments_concatenated.append({"speaker": speaker, "begin": segment.start, "end": segment.end})
-            else:
-                segments_concatenated[-1]['end'] = segment.end
+            base_time = self._chunk_index * self.chunk_duration_seconds + self.global_time_offset
+            current_spk = current_chunk_preds[0]
+            start_time = round(base_time, 2)
+            for idx, spk in enumerate(current_chunk_preds):
+                current_time = round(base_time + idx * frame_duration, 2)
+                if spk != current_spk:
+                    new_segments.append(SpeakerSegment(
+                        speaker=current_spk,
+                        start=start_time,
+                        end=current_time
+                    ))
+                    start_time = current_time
+                    current_spk = spk
+            new_segments.append(
+                SpeakerSegment(
+                        speaker=current_spk,
+                        start=start_time,
+                        end=current_time
+            )
+            )
+        return new_segments
                 
-        return segments_concatenated
-
     def get_segments(self) -> List[SpeakerSegment]:
         """Get a copy of the current speaker segments."""
         with self.segment_lock:
-            return self.speaker_segments.copy()
-
-    def clear_old_segments(self, older_than: float = 30.0):
-        """Clear segments older than the specified time."""
-        with self.segment_lock:
-            current_time = self.processed_time
-            self.speaker_segments = [
-                segment for segment in self.speaker_segments 
-                if current_time - segment.end < older_than
-            ]
-            logger.debug(f"Cleared old segments, remaining: {len(self.speaker_segments)}")
+            return self.diarization_segments.copy()
 
     def close(self):
         """Close the diarization system and clean up resources."""
         logger.info("Closing SortformerDiarization")
         with self.segment_lock:
-            self.speaker_segments.clear()
+            self.diarization_segments.clear()
         
         if self.debug:
             concatenated_audio = np.concatenate(self.audio_buffer)
@@ -434,11 +296,12 @@ def extract_number(s: str) -> int:
 
 if __name__ == '__main__':
     import asyncio
+
     import librosa
     
     async def main():
         """TEST ONLY."""
-        an4_audio = 'audio_test.mp3'
+        an4_audio = 'diarization_audio.wav'
         signal, sr = librosa.load(an4_audio, sr=16000)
         signal = signal[:16000*30]
 
@@ -450,13 +313,15 @@ if __name__ == '__main__':
         print("Speaker 0: 0:25 - 0:30")
         print("=" * 50)
         
-        diarization = SortformerDiarization(sample_rate=16000)        
+        diarization_backend = SortformerDiarization()
+        diarization = SortformerDiarizationOnline(shared_model = diarization_backend)     
         chunk_size = 1600
         
         for i in range(0, len(signal), chunk_size):
             chunk = signal[i:i+chunk_size]
-            await diarization.diarize(chunk)
+            new_segments = await diarization.diarize(chunk)
             print(f"Processed chunk {i // chunk_size + 1}")
+            print(new_segments)
         
         segments = diarization.get_segments()
         print("\nDiarization results:")

whisperlivekit/diarization/sortformer_backend_offline.py

@@ -1,205 +0,0 @@
-import numpy as np
-import torch
-import logging
-
-from nemo.collections.asr.models import SortformerEncLabelModel
-from nemo.collections.asr.modules import AudioToMelSpectrogramPreprocessor
-import librosa
-
-logger = logging.getLogger(__name__)
-
-def load_model():
-
-    diar_model = SortformerEncLabelModel.from_pretrained("nvidia/diar_streaming_sortformer_4spk-v2")
-    diar_model.eval()
-
-    if torch.cuda.is_available():
-        diar_model.to(torch.device("cuda"))
-
-    #we target 1 second lag for the moment. chunk_len could be reduced.
-    diar_model.sortformer_modules.chunk_len = 10
-    diar_model.sortformer_modules.subsampling_factor = 10 #8 would be better ideally
-
-    diar_model.sortformer_modules.chunk_right_context = 0 #no.
-    diar_model.sortformer_modules.chunk_left_context = 10 #big so it compensiate the problem with no padding later.
-
-    diar_model.sortformer_modules.spkcache_len = 188
-    diar_model.sortformer_modules.fifo_len = 188
-    diar_model.sortformer_modules.spkcache_update_period = 144
-    diar_model.sortformer_modules.log = False
-    diar_model.sortformer_modules._check_streaming_parameters()
-
-
-    audio2mel = AudioToMelSpectrogramPreprocessor(
-            window_size= 0.025, 
-            normalize="NA",
-            n_fft=512,
-            features=128,
-            pad_to=0) #pad_to 16 works better than 0. On test audio, we detect a third speaker for 1 second with pad_to=0. To solve that : increase left context to 10.
-
-    return diar_model, audio2mel
-
-diar_model, audio2mel = load_model()
-
-class StreamingSortformerState:
-    """
-    This class creates a class instance that will be used to store the state of the
-    streaming Sortformer model.
-
-    Attributes:
-        spkcache (torch.Tensor): Speaker cache to store embeddings from start
-        spkcache_lengths (torch.Tensor): Lengths of the speaker cache
-        spkcache_preds (torch.Tensor): The speaker predictions for the speaker cache parts
-        fifo (torch.Tensor): FIFO queue to save the embedding from the latest chunks
-        fifo_lengths (torch.Tensor): Lengths of the FIFO queue
-        fifo_preds (torch.Tensor): The speaker predictions for the FIFO queue parts
-        spk_perm (torch.Tensor): Speaker permutation information for the speaker cache
-        mean_sil_emb (torch.Tensor): Mean silence embedding
-        n_sil_frames (torch.Tensor): Number of silence frames
-    """
-
-    spkcache = None  # Speaker cache to store embeddings from start
-    spkcache_lengths = None  #
-    spkcache_preds = None  # speaker cache predictions
-    fifo = None  # to save the embedding from the latest chunks
-    fifo_lengths = None
-    fifo_preds = None
-    spk_perm = None
-    mean_sil_emb = None
-    n_sil_frames = None
-
-
-def init_streaming_state(self, batch_size: int = 1, async_streaming: bool = False, device: torch.device = None):
-    """
-    Initializes StreamingSortformerState with empty tensors or zero-valued tensors.
-
-    Args:
-        batch_size (int): Batch size for tensors in streaming state
-        async_streaming (bool): True for asynchronous update, False for synchronous update
-        device (torch.device): Device for tensors in streaming state
-
-    Returns:
-        streaming_state (SortformerStreamingState): initialized streaming state
-    """
-    streaming_state = StreamingSortformerState()
-    if async_streaming:
-        streaming_state.spkcache = torch.zeros((batch_size, self.spkcache_len, self.fc_d_model), device=device)
-        streaming_state.spkcache_preds = torch.zeros((batch_size, self.spkcache_len, self.n_spk), device=device)
-        streaming_state.spkcache_lengths = torch.zeros((batch_size,), dtype=torch.long, device=device)
-        streaming_state.fifo = torch.zeros((batch_size, self.fifo_len, self.fc_d_model), device=device)
-        streaming_state.fifo_lengths = torch.zeros((batch_size,), dtype=torch.long, device=device)
-    else:
-        streaming_state.spkcache = torch.zeros((batch_size, 0, self.fc_d_model), device=device)
-        streaming_state.fifo = torch.zeros((batch_size, 0, self.fc_d_model), device=device)
-    streaming_state.mean_sil_emb = torch.zeros((batch_size, self.fc_d_model), device=device)
-    streaming_state.n_sil_frames = torch.zeros((batch_size,), dtype=torch.long, device=device)
-    return streaming_state
-
-
-def process_diarization(chunks):
-    """ 
-    what it does:
-    1. Preprocessing: Applies dithering and pre-emphasis (high-pass filter) if enabled
-    2. STFT: Computes the Short-Time Fourier Transform using:
-        - the window of window_size=0.025 --> size of a window : 400 samples
-        - the hop parameter : n_window_stride = 0.01 -> every 160 samples, a new window
-    3. Magnitude Calculation: Converts complex STFT output to magnitude spectrogram
-    4. Mel Conversion: Applies Mel filterbanks (128 filters in this case) to get Mel spectrogram
-    5. Logarithm: Takes the log of the Mel spectrogram (if `log=True`)
-    6. Normalization: Skips normalization since `normalize="NA"`
-    7. Padding: Pads the time dimension to a multiple of `pad_to` (default 16)    
-    """
-    previous_chunk = None
-    l_chunk_feat_seq_t = []
-    for chunk in chunks:
-        audio_signal_chunk = torch.tensor(chunk).unsqueeze(0).to(diar_model.device)
-        audio_signal_length_chunk = torch.tensor([audio_signal_chunk.shape[1]]).to(diar_model.device)
-        processed_signal_chunk, processed_signal_length_chunk = audio2mel.get_features(audio_signal_chunk, audio_signal_length_chunk)
-        if previous_chunk is not None:
-            to_add = previous_chunk[:, :, -99:]
-            total = torch.concat([to_add, processed_signal_chunk], dim=2)
-        else:
-            total = processed_signal_chunk
-        previous_chunk = processed_signal_chunk
-        l_chunk_feat_seq_t.append(torch.transpose(total, 1, 2))
-
-    batch_size = 1
-    streaming_state = init_streaming_state(diar_model.sortformer_modules,
-        batch_size = batch_size,
-        async_streaming = True,
-        device = diar_model.device
-    )
-    total_preds = torch.zeros((batch_size, 0, diar_model.sortformer_modules.n_spk), device=diar_model.device)
-
-    chunk_duration_seconds = diar_model.sortformer_modules.chunk_len * diar_model.sortformer_modules.subsampling_factor * diar_model.preprocessor._cfg.window_stride
-
-    l_speakers = [
-        {'start_time': 0,
-        'end_time': 0,
-        'speaker': 0
-        }
-    ]
-    len_prediction = None
-    left_offset = 0
-    right_offset = 8
-    for i, chunk_feat_seq_t in enumerate(l_chunk_feat_seq_t):
-        with torch.inference_mode():
-                streaming_state, total_preds = diar_model.forward_streaming_step(
-                    processed_signal=chunk_feat_seq_t,
-                    processed_signal_length=torch.tensor([chunk_feat_seq_t.shape[1]]),
-                    streaming_state=streaming_state,
-                    total_preds=total_preds,
-                    left_offset=left_offset,
-                    right_offset=right_offset,
-                )
-                left_offset = 8
-                preds_np = total_preds[0].cpu().numpy()
-                active_speakers = np.argmax(preds_np, axis=1)
-                if len_prediction is None:
-                    len_prediction = len(active_speakers) # we want to get the len of 1 prediction
-                frame_duration = chunk_duration_seconds / len_prediction
-                active_speakers = active_speakers[-len_prediction:]
-                for idx, spk in enumerate(active_speakers):
-                    if spk != l_speakers[-1]['speaker']:
-                        l_speakers.append(
-                            {'start_time': (i * chunk_duration_seconds + idx * frame_duration),
-                            'end_time': (i * chunk_duration_seconds + (idx + 1) * frame_duration),
-                            'speaker': spk
-                        })                    
-                    else:
-                        l_speakers[-1]['end_time'] = i * chunk_duration_seconds + (idx + 1) * frame_duration
-                    
-        
-        """
-        Should print
-        [{'start_time': 0, 'end_time': 8.72, 'speaker': 0}, 
-        {'start_time': 8.72, 'end_time': 18.88, 'speaker': 1},
-        {'start_time': 18.88, 'end_time': 24.96, 'speaker': 2},
-        {'start_time': 24.96, 'end_time': 31.68, 'speaker': 0}]
-        """
-    for speaker in l_speakers:
-        print(f"Speaker {speaker['speaker']}: {speaker['start_time']:.2f}s - {speaker['end_time']:.2f}s")    
-    
-
-if __name__ == '__main__':
-    
-    an4_audio = 'audio_test.mp3'
-    signal, sr = librosa.load(an4_audio, sr=16000)
-    signal = signal[:16000*30]
-    # signal = signal[:-(len(signal)%16000)]
-
-    print("\n" + "=" * 50)
-    print("Expected ground truth:")
-    print("Speaker 0: 0:00 - 0:09")
-    print("Speaker 1: 0:09 - 0:19") 
-    print("Speaker 2: 0:19 - 0:25")
-    print("Speaker 0: 0:25 - 0:30")
-    print("=" * 50)
-
-    chunk_size = 16000  # 1 second
-    chunks = []
-    for i in range(0, len(signal), chunk_size):
-        chunk = signal[i:i+chunk_size]
-        chunks.append(chunk)
-    
-    process_diarization(chunks)

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

@@ -1,21 +1,25 @@
-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:
@@ -44,24 +48,23 @@ 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 +168,8 @@ 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)
@@ -224,7 +227,8 @@ class MLXWhisper(ASRBase):
             if segment.get("no_speech_prob", 0) > 0.9:
                 continue
             for word in segment.get("words", []):
-                token = ASRToken(word["start"], word["end"], word["word"], probability=word["probability"])
+                probability=word["probability"]
+                token = ASRToken(word["start"], word["end"], word["word"])
                 tokens.append(token)
         return tokens
 
@@ -245,6 +249,7 @@ 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
@@ -290,7 +295,8 @@ 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__)
@@ -151,21 +153,32 @@ class OnlineASRProcessor:
         """Append an audio chunk (a numpy array) to the current audio buffer."""
         self.audio_buffer = np.append(self.audio_buffer, audio)
 
-    def insert_silence(self, silence_duration, offset):
-        """
-        If silences are > 5s, we do a complete context clear. Otherwise, we just insert a small silence and shift the last_attend_frame
-        """
-        # if self.transcript_buffer.buffer:
-        #     self.committed.extend(self.transcript_buffer.buffer)
-        #     self.transcript_buffer.buffer = []
-            
-        if True: #silence_duration < 3: #we want the last audio to be treated to not have a gap. could also be handled in the future in ends_with_silence.
-            gap_silence = np.zeros(int(16000 * silence_duration), dtype=np.int16)
-            self.insert_audio_chunk(gap_silence)
+    def start_silence(self):
+        if self.audio_buffer.size == 0:
+            return [], self.get_audio_buffer_end_time()
+        return self.process_iter()
+
+    def end_silence(self, silence_duration: Optional[float], offset: float):
+        if not silence_duration or silence_duration <= 0:
+            return
+
+        long_silence = silence_duration >= 5
+        if not long_silence:
+            gap_samples = int(self.SAMPLING_RATE * silence_duration)
+            if gap_samples > 0:
+                gap_silence = np.zeros(gap_samples, dtype=np.float32)
+                self.insert_audio_chunk(gap_silence)
         else:
             self.init(offset=silence_duration + offset)
+
         self.global_time_offset += silence_duration
 
+    def insert_silence(self, silence_duration, offset):
+        """
+        Backwards compatibility shim for legacy callers that still use insert_silence.
+        """
+        self.end_silence(silence_duration, offset)
+
     def prompt(self) -> Tuple[str, str]:
         """
         Returns a tuple: (prompt, context), where:
@@ -400,11 +413,11 @@ class OnlineASRProcessor:
     ) -> Transcript:
         sep = sep if sep is not None else self.asr.sep
         text = sep.join(token.text for token in tokens)
-        probability = sum(token.probability for token in tokens if token.probability) / len(tokens) if tokens else None
+        # probability = sum(token.probability for token in tokens if token.probability) / len(tokens) if tokens else None
         if tokens:
             start = offset + tokens[0].start
             end = offset + tokens[-1].end
         else:
             start = None
             end = None
-        return Transcript(start, end, text, probability=probability)
+        return Transcript(start, end, text)

whisperlivekit/local_agreement/whisper_online.py

@@ -1,18 +1,19 @@
 #!/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 sys
+import time
+from functools import lru_cache
+
+import librosa
+import numpy as np
+
+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 +77,7 @@ def backend_factory(
             model_cache_dir,
             model_dir,
             model_path,
+            lora_path,
             direct_english_translation,
             buffer_trimming,
             buffer_trimming_sec,
@@ -86,16 +88,20 @@ 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 +126,8 @@ 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 +139,14 @@ 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 +155,9 @@ 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/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(
@@ -81,14 +82,14 @@ def parse_args():
     parser.add_argument(
         "--min-chunk-size",
         type=float,
-        default=0.5,
+        default=0.1,
         help="Minimum audio chunk size in seconds. It waits up to this time to do processing. If the processing takes shorter time, it waits, otherwise it processes the whole segment that was received by this time.",
     )
     
     parser.add_argument(
         "--model",
         type=str,
-        default="small",
+        default="base",
         dest='model_size',
         help="Name size of the Whisper model to use (default: tiny). Suggested values: tiny.en,tiny,base.en,base,small.en,small,medium.en,medium,large-v1,large-v2,large-v3,large,large-v3-turbo. The model is automatically downloaded from the model hub if not present in model cache dir.",
     )
@@ -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-mlx"],
+        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. Use 'voxtral-mlx' for Voxtral streaming 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,

whisperlivekit/remove_silences.py

@@ -1,106 +0,0 @@
-from whisperlivekit.timed_objects import ASRToken
-from time import time
-import re
-
-MIN_SILENCE_DURATION = 4 #in seconds
-END_SILENCE_DURATION = 8 #in seconds. you should keep it important to not have false positive when the model lag is important
-END_SILENCE_DURATION_VAC = 3 #VAC is good at detecting silences, but we want to skip the smallest silences
-
-def blank_to_silence(tokens):
-    full_string = ''.join([t.text for t in tokens])
-    patterns = [re.compile(r'(?:\s*\[BLANK_AUDIO\]\s*)+'), re.compile(r'(?:\s*\[typing\]\s*)+')]             
-    matches = []
-    for pattern in patterns:
-        for m in pattern.finditer(full_string):
-            matches.append({
-                'start': m.start(),
-                'end': m.end()
-            })
-    if matches:
-        # cleaned = pattern.sub(' ', full_string).strip()
-        # print("Cleaned:", cleaned)
-        cumulated_len = 0
-        silence_token = None
-        cleaned_tokens = []
-        for token in tokens:
-            if matches:
-                start = cumulated_len
-                end = cumulated_len + len(token.text)
-                cumulated_len = end
-                if start >= matches[0]['start'] and end <= matches[0]['end']:
-                    if silence_token: #previous token was already silence
-                        silence_token.start = min(silence_token.start, token.start)
-                        silence_token.end = max(silence_token.end, token.end)
-                    else: #new silence
-                        silence_token = ASRToken(
-                            start=token.start,
-                            end=token.end,
-                            speaker=-2,
-                            probability=0.95
-                        )
-                else:
-                    if silence_token: #there was silence but no more
-                        if silence_token.duration() >= MIN_SILENCE_DURATION:
-                            cleaned_tokens.append(
-                                silence_token
-                            )
-                        silence_token = None
-                        matches.pop(0)
-                    cleaned_tokens.append(token)
-        # print(cleaned_tokens)    
-        return cleaned_tokens
-    return tokens
-
-def no_token_to_silence(tokens):
-    new_tokens = []
-    silence_token = None
-    for token in tokens:
-        if token.speaker == -2:
-            if new_tokens and new_tokens[-1].speaker == -2: #if token is silence and previous one too
-                new_tokens[-1].end = token.end
-            else:
-                new_tokens.append(token)
-            
-        last_end = new_tokens[-1].end if new_tokens else 0.0
-        if token.start - last_end >= MIN_SILENCE_DURATION: #if token is not silence but important gap
-            if new_tokens and new_tokens[-1].speaker == -2:
-                new_tokens[-1].end = token.start
-            else:
-                silence_token = ASRToken(
-                    start=last_end,
-                    end=token.start,
-                    speaker=-2,
-                    probability=0.95
-                    )
-                new_tokens.append(silence_token)
-        
-        if token.speaker != -2:
-            new_tokens.append(token)
-    return new_tokens
-            
-def ends_with_silence(tokens, beg_loop, vac_detected_silence):
-    current_time = time() - (beg_loop if beg_loop else 0.0)
-    last_token = tokens[-1]
-    if  vac_detected_silence or (current_time - last_token.end >= END_SILENCE_DURATION):
-        if last_token.speaker == -2:
-            last_token.end = current_time
-        else:
-            tokens.append(
-                ASRToken(
-                    start=tokens[-1].end,
-                    end=current_time,
-                    speaker=-2,
-                    probability=0.95
-                )
-            )
-    return tokens
-    
-
-def handle_silences(tokens, beg_loop, vac_detected_silence):
-    if not tokens:
-        return []
-    tokens = blank_to_silence(tokens) #useful for simulstreaming backend which tends to generate [BLANK_AUDIO] text
-    tokens = no_token_to_silence(tokens)
-    tokens = ends_with_silence(tokens, beg_loop, vac_detected_silence)
-    return tokens
-     

whisperlivekit/results_formater.py

@@ -1,153 +0,0 @@
-
-import logging
-from whisperlivekit.remove_silences import handle_silences
-from whisperlivekit.timed_objects import Line, format_time
-
-logger = logging.getLogger(__name__)
-logger.setLevel(logging.DEBUG)
-
-CHECK_AROUND = 4
-DEBUG = False
-
-
-def is_punctuation(token):
-    if token.is_punctuation():
-        return True
-    return False
-
-def next_punctuation_change(i, tokens):
-    for ind in range(i+1, min(len(tokens), i+CHECK_AROUND+1)):
-        if is_punctuation(tokens[ind]):
-            return ind        
-    return None
-
-def next_speaker_change(i, tokens, speaker):
-    for ind in range(i-1, max(0, i-CHECK_AROUND)-1, -1):
-        token = tokens[ind]
-        if is_punctuation(token):
-            break
-        if token.speaker != speaker:
-            return ind, token.speaker
-    return None, speaker
-    
-def new_line(
-    token,
-):
-    return Line(
-        speaker = token.corrected_speaker,
-        text = token.text + (f"[{format_time(token.start)} : {format_time(token.end)}]" if DEBUG else ""),
-        start = token.start,
-        end = token.end,
-        detected_language=token.detected_language
-    )
-
-def append_token_to_last_line(lines, sep, token):
-    if not lines:
-        lines.append(new_line(token))
-    else:
-        if token.text:
-            lines[-1].text += sep + token.text + (f"[{format_time(token.start)} : {format_time(token.end)}]" if DEBUG else "")
-            lines[-1].end = token.end
-        if not lines[-1].detected_language and token.detected_language:
-            lines[-1].detected_language = token.detected_language
-            
-
-def format_output(state, silence, args, sep):
-    diarization = args.diarization
-    disable_punctuation_split = args.disable_punctuation_split
-    tokens = state.tokens
-    translation_validated_segments = state.translation_validated_segments # Here we will attribute the speakers only based on the timestamps of the segments
-    last_validated_token = state.last_validated_token
-    
-    previous_speaker = 1
-    undiarized_text = []
-    tokens = handle_silences(tokens, state.beg_loop, silence)
-    for i in range(last_validated_token, len(tokens)):
-        token = tokens[i]
-        speaker = int(token.speaker)
-        token.corrected_speaker = speaker
-        if not diarization:
-            if speaker == -1: #Speaker -1 means no attributed by diarization. In the frontend, it should appear under 'Speaker 1'
-                token.corrected_speaker = 1
-                token.validated_speaker = True
-        else:
-                if is_punctuation(token):
-                    state.last_punctuation_index = i  
-                
-                if state.last_punctuation_index == i-1:
-                    if token.speaker != previous_speaker:
-                        token.validated_speaker = True
-                        # perfect, diarization perfectly aligned
-                        last_punctuation = None
-                    else:
-                        speaker_change_pos, new_speaker = next_speaker_change(i, tokens, speaker)
-                        if speaker_change_pos:
-                            # Corrects delay:
-                            # That was the idea. <Okay> haha |SPLIT SPEAKER| that's a good one 
-                            # should become:
-                            # That was the idea. |SPLIT SPEAKER| <Okay> haha that's a good one 
-                            token.corrected_speaker = new_speaker
-                            token.validated_speaker = True
-                elif speaker != previous_speaker:
-                    if not (speaker == -2 or previous_speaker == -2):
-                        if next_punctuation_change(i, tokens):
-                            # Corrects advance:
-                            # Are you |SPLIT SPEAKER| <okay>? yeah, sure. Absolutely 
-                            # should become:
-                            # Are you <okay>? |SPLIT SPEAKER| yeah, sure. Absolutely 
-                            token.corrected_speaker = previous_speaker
-                            token.validated_speaker = True
-                        else: #Problematic, except if the language has no punctuation. We append to previous line, except if disable_punctuation_split is set to True.
-                            if not disable_punctuation_split:
-                                token.corrected_speaker = previous_speaker
-                                token.validated_speaker = False
-        if token.validated_speaker:
-            state.last_validated_token = i
-        previous_speaker = token.corrected_speaker  
-
-    previous_speaker = 1
-    
-    lines = []
-    for token in tokens:
-        if int(token.corrected_speaker) != int(previous_speaker):
-            lines.append(new_line(token))
-        else:
-            append_token_to_last_line(lines, sep, token)
-
-        previous_speaker = token.corrected_speaker        
-
-    if lines:
-        unassigned_translated_segments = []
-        for ts in translation_validated_segments:
-            assigned = False
-            for line in lines:
-                if ts and ts.overlaps_with(line):
-                    if ts.is_within(line):
-                        line.translation += ts.text + ' '
-                        assigned = True
-                        break
-                    else:
-                        ts0, ts1 = ts.approximate_cut_at(line.end)
-                        if ts0 and line.overlaps_with(ts0):
-                            line.translation += ts0.text + ' '
-                        if ts1:
-                            unassigned_translated_segments.append(ts1)
-                        assigned = True
-                        break
-            if not assigned:
-                unassigned_translated_segments.append(ts)
-        
-        if unassigned_translated_segments:
-            for line in lines:
-                remaining_segments = []
-                for ts in unassigned_translated_segments:
-                    if ts and ts.overlaps_with(line):
-                        line.translation += ts.text + ' '
-                    else:
-                        remaining_segments.append(ts)
-                unassigned_translated_segments = remaining_segments #maybe do smth in the future about that
-    
-    if state.buffer_transcription and lines:
-        lines[-1].end = max(state.buffer_transcription.end, lines[-1].end)
-        
-    return lines, undiarized_text

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)
@@ -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:
+    if opset_version not in available_ops:
         raise Exception(f'Available ONNX opset_version: {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
 
 
@@ -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/backend.py

@@ -1,96 +1,104 @@
-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 PaddedAlignAttWhisper
 
 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
 else:
     WhisperModel = None
 
+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 = PaddedAlignAttWhisper(
-            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 insert_silence(self, silence_duration, offset):
-        """
-        If silences are > 5s, we do a complete context clear. Otherwise, we just insert a small silence and shift the last_attend_frame
-        """
-        if silence_duration < 5:
-            gap_silence = torch.zeros(int(16000*silence_duration))
-            self.model.insert_audio(gap_silence)
-            # self.global_time_offset += silence_duration
-        else:
-            self.process_iter(is_last=True) #we want to totally process what remains in the buffer.
+    def start_silence(self):
+        tokens, processed_upto = self.process_iter(is_last=True)
+        return tokens, processed_upto
+
+    def end_silence(self, silence_duration, offset):
+        """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:
+                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='')
@@ -104,15 +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
@@ -120,6 +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)
@@ -128,14 +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 = ""
 
@@ -152,35 +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.")
@@ -195,6 +199,10 @@ class SimulStreamingASR():
         self.fast_encoder = self.encoder_backend in ("mlx-whisper", "faster-whisper")
         if self.encoder_backend == "whisper":
             self.disable_fast_encoder = True
+        
+        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,
@@ -206,7 +214,7 @@ 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,
@@ -218,24 +226,37 @@ 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:
@@ -245,8 +266,20 @@ 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):
@@ -290,44 +323,29 @@ 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:                
-                temp_model = PaddedAlignAttWhisper(
+            if self.fast_encoder:
+                temp_model = AlignAtt(
                     cfg=self.cfg,
                     loaded_model=whisper_model,
                     mlx_encoder=self.mlx_encoder,
                     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."""

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 +1,95 @@
+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)
+    
+    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.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/generation_progress.py

@@ -1,43 +0,0 @@
-class Tokens:
-    def __init__(self, tokens):
-        self.tokens = tokens
-
-#    def clone(self):
-#        return Tokens(self.tokens.clone())
-
-    def __str__(self):
-        return str(self.tokens.tolist())
-
-    def __repr__(self):
-        return self.__str__()
-
-class BeamTokens(Tokens):
-    def __init__(self, tokens, beam_size):
-        self.tokens = tokens
-        self.beam_size = beam_size
-
-    def clone(self):
-        return BeamTokens(self.tokens.clone())
-
-    def __str__(self):
-        return f"BeamTokens({self.tokens.tolist()}, beam_size={self.beam_size})"
-
-    def __repr__(self):
-        return self.__str__()
-
-    def as_text(self, tokenizer):
-        return tokenizer.decode(self.tokens)
-
-class Logits(Tokens):
-    def __init__(self, logits):
-        super().__init__(logits)
-
-#    def clone(self):
-#        return Logits(self.tokens.clone(), self.beam_size)
-
-    def __str__(self):
-#        return "abc"
-        return f"Logits({self.tokens.shape})"
-
-    def __repr__(self):
-        return self.__str__()

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 +1,76 @@
+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)
+    
+    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.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 +1,756 @@
+"""
+MLX whisper AlignAtt streaming decoder
+"""
+import logging
+from time import time
+from typing import Any, List, Optional, 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.timed_objects import ASRToken
+from whisperlivekit.whisper import DecodingOptions, tokenizer
+from whisperlivekit.whisper.audio import N_FRAMES, N_SAMPLES, TOKENS_PER_SECOND
+
+from ..config import AlignAttConfig
+from .decoder_state import MLXDecoderState
+from .decoders import MLXBeamSearchDecoder, MLXGreedyDecoder, MLXInference
+
+DEC_PAD = 50257
+logger = logging.getLogger(__name__)
+
+
+class MLXTokenBuffer: #should try to make it heritate from classic simul whisper class
+    """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:
+        """Return tokens as MLX 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:
+        """Return tokens as MLX array repeated for beam search."""
+        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):
+        """Trim words from the beginning of the context."""
+        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):
+        """Append token IDs to the buffer, handling incomplete UTF-8."""
+        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.
+    
+    Args:
+        x: Input array of shape (..., T)
+        filter_width: Width of the median filter (should be odd)
+        
+    Returns:
+        Filtered array of same shape
+    """
+    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_shape = list(shape)
+    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:
+    """
+    MLX-native Alignment-based Attention decoder for SimulStreaming.
+    
+    This class runs entirely on MLX, with no PyTorch dependencies for inference.
+    """
+
+    @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
+
+    def __init__(
+        self,
+        cfg: AlignAttConfig,
+        mlx_model: Any,
+    ) -> None:
+        """
+        Initialize MLX AlignAtt decoder.
+        
+        Args:
+            cfg: AlignAtt configuration
+            mlx_model: MLX Whisper model (full model, not just encoder)
+        """
+        self.model = mlx_model
+        self.cfg = cfg
+        
+        logger.info(f"MLX Model dimensions: {self.model.dims}")
+        
+        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 = self.model.dims.n_text_ctx
+        self.num_decoder_layers = len(self.model.decoder.blocks)
+
+        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
+
+        # Initialize per-session state
+        self.state = MLXDecoderState()
+        self._init_state(cfg)
+
+    def _init_state(self, cfg: AlignAttConfig):
+        """Initialize the per-session decoder 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
+
+        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 = [
+            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()
+
+        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
+
+    def warmup(self, audio: np.ndarray):
+        """Warmup the model with sample audio."""
+        try:
+            self.insert_audio(audio)
+            self.infer(is_last=True)
+            self.refresh_segment(complete=True)
+            logger.info("MLX model warmed up successfully")
+        except Exception as e:
+            logger.exception(f"MLX model warmup failed: {e}")
+
+    def create_tokenizer(self, language=None):
+        """Create tokenizer for the given language."""
+        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 init_context(self):
+        """Initialize context buffer."""
+        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 init_tokens(self):
+        """Initialize token sequence."""
+        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 trim_context(self):
+        """Trim context if too long."""
+        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
+        if self.cfg.static_init_prompt is None:
+            after = 0
+        else:
+            after = 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):
+        """Refresh segment state."""
+        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 = []
+
+    def fire_at_boundary(self, chunked_encoder_feature: mx.array) -> bool:
+        """Check if we should fire at word boundary (CIF-based)."""
+        if self.state.always_fire:
+            return True
+        if self.state.never_fire:
+            return False
+        return True
+
+    def _current_tokens(self) -> mx.array:
+        """Get current token sequence for decoding."""
+        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
+
+        # Concatenate all tokens
+        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 debug_print_tokens(self, tokens: mx.array):
+        """Debug print token sequences."""
+        tokens_np = np.array(tokens)
+        for i in range(min(self.cfg.beam_size, tokens_np.shape[0])):
+            logger.debug(self.tokenizer.decode_with_timestamps(tokens_np[i].tolist()))
+
+    def segments_len(self) -> float:
+        """Get total length of audio segments in seconds."""
+        return sum(s.shape[0] for s in self.state.segments) / 16000
+
+    def _apply_minseglen(self) -> bool:
+        """Check if we have enough audio to process."""
+        segments_len = self.segments_len()
+        if segments_len < self.cfg.audio_min_len:
+            logger.debug("waiting for next segment")
+            return False
+        return True
+
+    def insert_audio(self, segment: np.ndarray = None):
+        """Insert audio segment into buffer."""
+        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)}, cumulative offset: {self.state.cumulative_time_offset:.2f}s")
+            
+            if len(self.state.tokens) > 1:
+                # Convert MLX array to list for context
+                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 _clean_cache(self):
+        """Clean the kv_cache after each inference step."""
+        self.state.clean_cache()
+
+    def _suppress_tokens(self, logits: mx.array) -> mx.array:
+        """Apply token suppression to 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 lang_id(self, encoder_features: mx.array) -> Tuple[mx.array, List[dict]]:
+        """Language detection from encoder features."""
+        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 infer(self, is_last: bool = False) -> List[ASRToken]:
+        """
+        Main inference method.
+        
+        Args:
+            is_last: Whether this is the final chunk
+            
+        Returns:
+            List of timestamped ASR tokens
+        """
+        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 []
+
+        if len(self.state.segments) > 1:
+            input_segments = np.concatenate(self.state.segments, axis=0)
+        else:
+            input_segments = self.state.segments[0]
+
+        beg_encode = time()
+        
+        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)
+        
+        mx.eval(encoder_feature)
+        
+        end_encode = time()
+        logger.debug(f'MLX Encoder duration: {end_encode - beg_encode:.3f}s')
+
+        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}")
+
+        self.trim_context()
+        current_tokens = self._current_tokens()
+
+        fire_detected = self.fire_at_boundary(encoder_feature[:, :content_mel_len, :])
+
+        sum_logprobs = mx.zeros((self.cfg.beam_size,), dtype=mx.float32)
+        completed = False
+
+        attn_of_alignment_heads = None
+        most_attended_frame = None
+
+        token_len_before_decoding = current_tokens.shape[1]
+
+        l_absolute_timestamps = []
+        accumulated_cross_attns = []
+
+        audio_duration_s = self.segments_len()
+        # ~15 text tokens/s is a generous upper bound for speech; TOKENS_PER_SECOND (50)
+        # is the mel-frame rate and was causing 10-40x over-allocation on repetition loops.
+        max_tokens_per_chunk = max(50, int(audio_duration_s * 15 * 1.5))
+        tokens_produced_this_chunk = 0
+
+        while not completed and current_tokens.shape[1] < self.max_text_len:
+            tokens_produced_this_chunk += 1
+
+            if tokens_produced_this_chunk > max_tokens_per_chunk:
+                logger.warning(f"[Loop Detection] Too many tokens ({tokens_produced_this_chunk}) for {audio_duration_s:.2f}s audio. Breaking.")
+                current_tokens = current_tokens[:, :token_len_before_decoding]
+                break
+
+            if new_segment:
+                tokens_for_logits = current_tokens
+            else:
+                tokens_for_logits = current_tokens[:, -1:]
+
+            if self.state.decoder_type == "greedy":
+                logits, self.state.kv_cache, cross_qk = self.model.decoder(
+                    tokens_for_logits, encoder_feature, kv_cache=self.state.kv_cache
+                )
+            else:
+                logits, cross_qk = self.state.inference.logits(tokens_for_logits, encoder_feature)
+
+            mx.eval(logits)
+            
+            accumulated_cross_attns.append(cross_qk)
+            if len(accumulated_cross_attns) > 16:
+                accumulated_cross_attns = accumulated_cross_attns[-16:]
+
+            if new_segment and 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")
+                    break
+
+            logits = logits[:, -1, :]  # Last token logits
+
+            # Suppress tokens at segment start
+            if new_segment:
+                blank_tokens = self.tokenizer.encode(" ") + [self.tokenizer.eot]
+                logits = logits.at[:, blank_tokens].add(-float('inf'))
+            new_segment = False
+            
+            logits = self._suppress_tokens(logits)
+            
+            current_tokens, completed = self.state.token_decoder.update(
+                current_tokens, logits, sum_logprobs
+            )
+            mx.eval(current_tokens)
+
+            logger.debug(f"Decoding completed: {completed}")
+            self.debug_print_tokens(current_tokens)
+
+            attn_of_alignment_heads = self._process_cross_attention(
+                accumulated_cross_attns, content_mel_len
+            )
+
+            most_attended_frames = mx.argmax(attn_of_alignment_heads[:, -1, :], axis=-1)
+            most_attended_frames_np = np.array(most_attended_frames)
+
+            absolute_timestamps = [
+                (frame * 0.02 + self.state.cumulative_time_offset)
+                for frame in most_attended_frames_np.tolist()
+            ]
+
+            logger.debug(str(most_attended_frames_np.tolist()) + " most att frames")
+            logger.debug(f"Absolute timestamps: {absolute_timestamps}")
+
+            most_attended_frame = int(most_attended_frames_np[0])
+            l_absolute_timestamps.append(absolute_timestamps[0])
+
+            if completed:
+                current_tokens = current_tokens[:, :-1]
+                break
+            if not is_last and self.state.last_attend_frame - most_attended_frame > self.cfg.rewind_threshold:
+                current_tokens_np = np.array(current_tokens)
+                if current_tokens.shape[1] > 1 and current_tokens_np[0, -2] >= DEC_PAD:
+                    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}, last: {self.state.last_attend_frame}")
+                    self.state.last_attend_frame = -self.cfg.rewind_threshold
+                    current_tokens = mx.concatenate(self.state.tokens, axis=1) if len(self.state.tokens) > 0 else self.state.tokens[0]
+                    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
+        tokens_to_split = np.array(current_tokens[0, token_len_before_decoding:]).tolist()
+        if self.state.pending_incomplete_tokens:
+            logger.debug(f"[UTF-8 Fix] Prepending pending tokens: {self.state.pending_incomplete_tokens}")
+            tokens_to_split = self.state.pending_incomplete_tokens + tokens_to_split
+
+        if fire_detected or is_last:
+            new_hypothesis = tokens_to_split
+            split_words, split_tokens = self.tokenizer.split_to_word_tokens(new_hypothesis)
+        else:
+            split_words, split_tokens = self.tokenizer.split_to_word_tokens(tokens_to_split)
+            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 = mx.array([new_hypothesis], dtype=mx.int32)
+        new_tokens = mx.repeat(new_tokens, self.cfg.beam_size, axis=0)
+        self.state.tokens.append(new_tokens)
+
+        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 = []
+        timestamp_idx = 0
+        replacement_char = "\ufffd"
+        
+        for word, word_tokens in zip(split_words, split_tokens):
+            if replacement_char in word:
+                logger.warning(f"[UTF-8 Filter] Skipping: {repr(word)}")
+                timestamp_idx += len(word_tokens)
+                continue
+
+            try:
+                current_timestamp = l_absolute_timestamps[timestamp_idx]
+            except IndexError:
+                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.state.speaker,
+                detected_language=self.state.detected_language
+            ).with_offset(self.state.global_time_offset)
+            timestamped_words.append(timestamp_entry)
+        self.state.pending_incomplete_tokens = []
+        MAX_PENDING_TOKENS = 10
+        if split_words and replacement_char in split_words[-1]:
+            if len(split_tokens[-1]) <= MAX_PENDING_TOKENS:
+                self.state.pending_incomplete_tokens = split_tokens[-1]
+                logger.debug(f"[UTF-8 Fix] Holding incomplete tokens")
+            else:
+                logger.warning(f"[UTF-8 Fix] Skipping too many tokens")
+
+        return timestamped_words
+
+    def _process_cross_attention(
+        self,
+        cross_attns: List[List[mx.array]],
+        content_mel_len: int
+    ) -> mx.array:
+        """
+        Process cross-attention weights for alignment.
+        
+        Args:
+            cross_attns: List of cross-attention from each forward pass
+                        Each element is a list of mx.arrays per layer
+            content_mel_len: Length of actual audio content
+            
+        Returns:
+            Processed attention tensor, shape (batch, seq_len, content_mel_len)
+        """
+        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 len(align_heads_in_layer) == 0:
+                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:
+                t = mx.concatenate(mat, axis=1)
+                tmp.append(t)
+
+        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
+

whisperlivekit/simul_whisper/mlx_encoder.py

@@ -5,7 +5,6 @@ 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 = {
@@ -69,4 +68,40 @@ 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

@@ -1,49 +1,84 @@
-# This code was originally in simul_whisper/transcriber/simul_whisper.py . It is adapted a lot for SimulStreaming.
-
-import os
 import logging
+import os
+from time import time
+from typing import List, Optional, Tuple
 
+import numpy as np
 import torch
 import torch.nn.functional as F
 
-from whisperlivekit.whisper import load_model, DecodingOptions, tokenizer
-from .config import AlignAttConfig
+from whisperlivekit.backend_support import (faster_backend_available,
+                                            mlx_backend_available)
 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.whisper import DecodingOptions, tokenizer
+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, detect_language
-from .beam import BeamPyTorchInference
-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,
-)
 
-import numpy as np
 from ..timed_objects import PUNCTUATION_MARKS
-from .generation_progress import *
+from .beam import BeamPyTorchInference
+from .config import AlignAttConfig
+from .decoder_state import DecoderState
+from .eow_detection import fire_at_boundary, load_cif
+from .token_buffer import TokenBuffer
 
 DEC_PAD = 50257
 logger = logging.getLogger(__name__)
 
-
-HAS_MLX_WHISPER = False
-HAS_FASTER_WHISPER = False
-
 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
-    HAS_MLX_WHISPER = True
 
 if faster_backend_available():
     from faster_whisper.audio import pad_or_trim as fw_pad_or_trim
     from faster_whisper.feature_extractor import FeatureExtractor
-    HAS_FASTER_WHISPER = True
 
-class PaddedAlignAttWhisper:
+USE_MLCORE = False
+
+
+def load_coreml_encoder():
+    try:
+        from coremltools.models import MLModel
+    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 = MLModel(COREML_ENCODER_PATH)
+    spec = _coreml_encoder.get_spec()
+    _coreml_input_name = spec.description.input[0].name if spec.description.input else "mel"
+    _coreml_output_name = spec.description.output[0].name if spec.description.output else None
+    return _coreml_encoder, _coreml_input_name, _coreml_output_name
+
+
+class AlignAtt:
+    """
+    Alignment-based Attention decoder for SimulStreaming.
+    
+    This class is now hookless - the model can be shared across multiple
+    sessions, with each session maintaining its own DecoderState.
+    """
+    
+    # Property accessors for backward compatibility
+    @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
+    
     def __init__(
             self, 
             cfg: AlignAttConfig,
@@ -51,130 +86,103 @@ class PaddedAlignAttWhisper:
             mlx_encoder=None,
             fw_encoder=None,
         ) -> None:
-        self.log_segments = 0
-        
+        # Shared model reference (can be shared across sessions)
         self.model = loaded_model
         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.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'
         
         logger.info(f"Model dimensions: {self.model.dims}")
-        self.speaker = -1
         self.decode_options = DecodingOptions(
-            language = cfg.language, 
-            without_timestamps = True,
+            language=cfg.language, 
+            without_timestamps=True,
             task=cfg.task
         )
         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
-        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
-
-
-        # tokens to be suppressed from decoding, to prevent hallucinations
-        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
-        if self.tokenizer.no_speech is not None:
-            suppress_tokens.append(self.tokenizer.no_speech)
-        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
-
-        # 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
+
+        # Initialize per-session state
+        self.state = DecoderState()
+        self._init_state(cfg)
+        
+    def _init_state(self, cfg: AlignAttConfig):
+        """Initialize the per-session decoder state."""
+        # Create tokenizer
+        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
+        
+        # Timing state
+        self.state.global_time_offset = 0.0
+        self.state.last_attend_frame = -cfg.rewind_threshold
+        self.state.speaker = -1
+        
+        # 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
+        )
+
+        # Build alignment source mapping from model's alignment_heads
+        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.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
+
+        # 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.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)))
+        logger.debug(f"Suppress tokens: {suppress_tokens}")
+        sup_tokens = SuppressTokens(suppress_tokens)
+        self.state.suppress_tokens_fn = lambda logits: sup_tokens.apply(logits, None)
+
+        # Initialize tokens
+        self.init_tokens()
         self.init_context()
 
-        # decoder type: greedy or beam
+        # Set up 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
-
-            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()
+            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
+            )
 
     def warmup(self, audio):
         try:
@@ -192,96 +200,100 @@ class PaddedAlignAttWhisper:
             num_languages=self.model.num_languages,
             task=self.decode_options.task
         )
+        self.state.tokenizer = self.tokenizer
 
     def init_context(self):
         kw = {'tokenizer': self.tokenizer, 
               'device': self.model.device, 
               'prefix_token_ids': [self.tokenizer.sot_prev]}
-        self.context = TokenBuffer.empty(**kw)
+        self.state.context = TokenBuffer.empty(**kw)
         if self.cfg.static_init_prompt is not None:
-            self.context = TokenBuffer.from_text(self.cfg.static_init_prompt, **kw)
+            self.state.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
+            self.state.context.text += self.cfg.init_prompt
 
     def init_tokens(self):
-        logger.debug(f"init tokens, {len(self.segments)}")
+        logger.debug(f"init tokens, {len(self.state.segments)}")
         # init tokens (mandatory prompt)
-        self.initial_tokens = torch.tensor(
+        self.state.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]
+        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}")
+        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
         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)
+            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.debug(f"len {l}, c {c}, max_context_tokens {self.max_context_tokens}")
-        logger.info(f"Context after trim: {self.context.text} (len: {l})")
+        logger.info(f"Context after trim: {self.state.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)
+    def logits(
+        self, 
+        tokens: torch.Tensor, 
+        audio_features: torch.Tensor,
+        return_cross_attn: bool = False
+    ):
+        """Get logits from decoder, optionally returning cross-attention weights."""
+        if self.state.decoder_type == "greedy":
+            return self.model.decoder(
+                tokens, audio_features, 
+                kv_cache=self.state.kv_cache,
+                return_cross_attn=return_cross_attn
+            )
         else:
             logger.debug(f"Logits shape: {tokens.shape}")
-            logit = self.inference.logits(tokens, audio_features)
-        return logit
+            return self.state.inference.logits(
+                tokens, audio_features,
+                return_cross_attn=return_cross_attn
+            )
     
 
     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.state.last_attend_frame = -self.cfg.rewind_threshold       
+        self.state.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:]
+        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.segments = []
-        self.log_segments += 1
-
-        self.pending_incomplete_tokens = []
+            self.state.segments = []
+        self.state.log_segments += 1
+        self.state.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)
-
+        if self.state.always_fire: 
+            return True
+        if self.state.never_fire: 
+            return False
+        return fire_at_boundary(chunked_encoder_feature, self.state.CIFLinear)
 
     def _current_tokens(self):
-
-        toks = self.tokens
+        toks = self.state.tokens
         # very first infer: duplicate start of seq to beam_size
         if toks[0].shape[0] == 1:
-            toks[0] = toks[0].repeat_interleave(self.cfg.beam_size,dim=0)
+            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)
+        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
@@ -301,7 +313,7 @@ class PaddedAlignAttWhisper:
     ### audio buffer 
 
     def segments_len(self):
-        segments_len = sum(s.shape[0] for s in self.segments) / 16000
+        segments_len = sum(s.shape[0] for s in self.state.segments) / 16000
         return segments_len
 
     def _apply_minseglen(self):
@@ -314,42 +326,36 @@ class PaddedAlignAttWhisper:
 
     def insert_audio(self, segment=None):
         if segment is not None:
-            self.segments.append(segment)
+            self.state.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
+        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.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:]
+            self.state.last_attend_frame -= int(TOKENS_PER_SECOND * removed_len)
+            self.state.cumulative_time_offset += removed_len  # Track cumulative time removed
+            self.state.segments = self.state.segments[1:]
+            logger.debug(f"remove segments: {len(self.state.segments)} {len(self.state.tokens)}, 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 _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()
+        """Clean the kv_cache after each inference step."""
+        self.state.clean_cache()
 
     @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 .
+        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]
+        # Note: don't use kv_cache for language detection
         logits = self.model.logits(x, encoder_features)[:, 0]
 
         # collect detected languages; suppress all non-language tokens
@@ -379,29 +385,41 @@ class PaddedAlignAttWhisper:
     @torch.no_grad()
     def infer(self, is_last=False):
         new_segment = True
-        if len(self.segments) == 0:
+        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()}.")
-            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)
+        if len(self.state.segments) > 1:
+            input_segments = torch.cat(self.state.segments, dim=0)
         else:
-            input_segments = self.segments[0]
+            input_segments = self.state.segments[0]
 
-        # if self.cfg.language == "auto" and self.reset_tokenizer_to_auto_next_call:
-        #     logger.debug("Resetting tokenizer to auto for new sentence.")
-        #     self.create_tokenizer(None)
-        #     self.detected_language = None
-        #     self.init_tokens()
-        #     self.reset_tokenizer_to_auto_next_call = False
-
-        # NEW : we can use a different encoder, before using standart whisper for cross attention with the hooks on the decoder
         beg_encode = time()
+        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",
+            ).unsqueeze(0)
+            mel = pad_or_trim(mel_padded, N_FRAMES)
+            content_mel_len = int((mel_padded.shape[2] - mel.shape[2]) / 2)
+            mel_np = np.ascontiguousarray(mel.numpy())
+            ml_inputs = {coreml_input_name or "mel": mel_np}
+            coreml_outputs = coreml_encoder.predict(ml_inputs)
+            if coreml_output_name and coreml_output_name in coreml_outputs:
+                encoder_feature_np = coreml_outputs[coreml_output_name]
+            else:
+                encoder_feature_np = next(iter(coreml_outputs.values()))
+            encoder_feature = torch.as_tensor(
+                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 = mlx_pad_or_trim(mlx_mel_padded, N_FRAMES, axis=-2)
@@ -432,18 +450,18 @@ class PaddedAlignAttWhisper:
         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 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.last_attend_frame = -self.cfg.rewind_threshold
-                self.cumulative_time_offset = 0.0
+                self.state.last_attend_frame = -self.cfg.rewind_threshold
+                self.state.cumulative_time_offset = 0.0
                 self.init_tokens()
                 self.init_context()
-                self.detected_language = top_lan
+                self.state.detected_language = top_lan
                 logger.info(f"Tokenizer language: {self.tokenizer.language}, {self.tokenizer.sot_sequence_including_notimestamps}")
 
         self.trim_context()
@@ -463,92 +481,96 @@ class PaddedAlignAttWhisper:
         
         l_absolute_timestamps = []
         
-        while not completed and current_tokens.shape[1] < self.max_text_len: # bos is 3 tokens
+        accumulated_cross_attns = []
+        
+        audio_duration_s = self.segments_len()
+        # ~15 text tokens/s is a generous upper bound for speech; TOKENS_PER_SECOND (50)
+        # is the mel-frame rate and was causing 10-40x over-allocation on repetition loops.
+        max_tokens_per_chunk = max(50, int(audio_duration_s * 15 * 1.5))
+        tokens_produced_this_chunk = 0
+        
+        while not completed and current_tokens.shape[1] < self.max_text_len:  # bos is 3 tokens
+            tokens_produced_this_chunk += 1
+            
+            if tokens_produced_this_chunk > max_tokens_per_chunk:
+                logger.warning(f"[Loop Detection] Too many tokens ({tokens_produced_this_chunk}) for {audio_duration_s:.2f}s audio. Breaking.")
+                current_tokens = current_tokens[:, :token_len_before_decoding]  # Discard all new tokens
+                break
 
             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:]
+                tokens_for_logits = current_tokens[:, -1:]
 
-            logits = self.logits(tokens_for_logits, encoder_feature) # B, len(tokens), token dict size
+            # Get logits and cross-attention weights from decoder
+            result = self.logits(tokens_for_logits, encoder_feature, return_cross_attn=True)
+            logits, cross_attns = result
+            
+            # Accumulate cross-attention from this forward pass (rolling window to
+            # bound VRAM — only the last entry matters for alignment, and the
+            # median_filter kernel is 7, so 16 entries is more than enough).
+            accumulated_cross_attns.append(cross_attns)
+            if len(accumulated_cross_attns) > 16:
+                accumulated_cross_attns = accumulated_cross_attns[-16:]
 
             if new_segment and self.tokenizer.no_speech is not None:
-                probs_at_sot = logits[:, self.sot_index, :].float().softmax(dim=-1)
+                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")
                     break
 
-            logits = logits[:, -1, :] # logits for the last token
+            logits = logits[:, -1, :]  # logits for the last token
 
-            # supress blank tokens only at the beginning of the segment
+            # suppress 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)
+            self.state.suppress_tokens_fn(logits)
+            current_tokens, completed = self.state.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]
+            # Process accumulated cross-attention weights for alignment
+            attn_of_alignment_heads = self._process_cross_attention(accumulated_cross_attns, content_mel_len)
 
             # for each beam, the most attended frame is:
-            most_attended_frames = torch.argmax(attn_of_alignment_heads[:,-1,:], dim=-1)
+            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()]
+            absolute_timestamps = [
+                (frame * 0.02 + self.state.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)")
+            logger.debug(f"Absolute timestamps: {absolute_timestamps} (offset: {self.state.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
+                # 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 not is_last and self.state.last_attend_frame - most_attended_frame > self.cfg.rewind_threshold:
                 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
+                    logger.debug("omit rewinding from special tokens")
+                    self.state.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]
+                        f"last attention pos: {self.state.last_attend_frame}; omit this segment")
+                    self.state.last_attend_frame = -self.cfg.rewind_threshold
+                    current_tokens = torch.cat(self.state.tokens, dim=1) if len(self.state.tokens) > 0 else self.state.tokens[0]
                     break
             else:
-                self.last_attend_frame = most_attended_frame
+                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}")
@@ -568,12 +590,12 @@ class PaddedAlignAttWhisper:
         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)
+        if self.state.pending_incomplete_tokens:
+            logger.debug(f"[UTF-8 Fix] Prepending {len(self.state.pending_incomplete_tokens)} pending tokens: {self.state.pending_incomplete_tokens}")
+            pending_tensor = torch.tensor(self.state.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:
+        if fire_detected or is_last:
             new_hypothesis = tokens_to_split.flatten().tolist()
             split_words, split_tokens = self.tokenizer.split_to_word_tokens(new_hypothesis)
         else:
@@ -584,20 +606,18 @@ class PaddedAlignAttWhisper:
             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)
+        self.state.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]
-            
+        if len(l_absolute_timestamps) >= 2 and self.state.first_timestamp is None:
+            self.state.first_timestamp = l_absolute_timestamps[0]
 
         timestamped_words = []
         timestamp_idx = 0
@@ -611,26 +631,96 @@ class PaddedAlignAttWhisper:
 
             try:
                 current_timestamp = l_absolute_timestamps[timestamp_idx]
-            except:
-                pass
+            except IndexError:
+                # Use last timestamp if index out of range
+                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=current_timestamp,
-                    end=current_timestamp + 0.1,
-                    text= word,
-                    probability=0.95,
-                    speaker=self.speaker,
-                    detected_language=self.detected_language
-                ).with_offset(
-                    self.global_time_offset
+                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)
 
-        # Hold incomplete tokens for next chunk
-        self.pending_incomplete_tokens = []
+        # Hold incomplete tokens for next chunk (with limit to prevent hallucination accumulation)
+        self.state.pending_incomplete_tokens = []
+        MAX_PENDING_TOKENS = 10  # Real incomplete UTF-8 chars are at most a few 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}")
+            if 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)} incomplete tokens for next chunk")
+            else:
+                logger.warning(f"[UTF-8 Fix] Skipping {len(split_tokens[-1])} tokens (exceeds limit of {MAX_PENDING_TOKENS}, likely hallucination)")
 
         return timestamped_words
+
+    def _process_cross_attention(
+        self, 
+        cross_attns: List[torch.Tensor], 
+        content_mel_len: int
+    ) -> torch.Tensor:
+        """
+        Process cross-attention weights from decoder layers for alignment.
+        
+        Args:
+            cross_attns: List of cross-attention tensors from each decoder layer.
+                         Each tensor has shape (batch, n_head, seq_len, audio_len)
+            content_mel_len: Length of actual audio content in mel frames
+            
+        Returns processed attention tensor for alignment, shape (batch, seq_len, content_mel_len)
+        """
+        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: List[torch.Tensor] = [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
+            # attn_mat shape: (batch, n_head, seq_len, audio_len) or (n_head, seq_len, audio_len) for batch=1
+            align_heads_in_layer = self.state.align_source.get(layer_rank, [])
+            if len(align_heads_in_layer) == 0:
+                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:
+                    # (n_head, seq_len, audio_len) when squeezed
+                    if attn_mat.dim() == 4:
+                        a = attn_mat[0, head_id, :, :]  # (seq_len, audio_len)
+                    else:
+                        a = attn_mat[head_id, :, :]
+                    a = a.unsqueeze(0)  # (1, seq_len, audio_len)
+                else:
+                    # attn_mat: (batch, n_head, seq_len, audio_len)
+                    a = attn_mat[:, head_id, :, :]  # (batch, seq_len, audio_len)
+                attn_of_alignment_heads[align_head_rank].append(a)
+        
+        tmp = []
+        for mat in attn_of_alignment_heads:
+            if mat:
+                t = torch.cat(mat, dim=1)  # (batch, total_seq_len, audio_len)
+                tmp.append(t)
+        
+        if not tmp:
+            return torch.zeros(self.cfg.beam_size, 1, content_mel_len, device=self.device)
+        
+        # stck al heads: (batch, num_align_heads, seq_len, audio_len)
+        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

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, Any, List
 from datetime import timedelta
+from typing import Any, Dict, List, Optional, Union
 
 PUNCTUATION_MARKS = {'.', '!', '?', '。', '！', '？'}
 
@@ -8,22 +8,19 @@ def format_time(seconds: float) -> str:
     """Format seconds as HH:MM:SS."""
     return str(timedelta(seconds=int(seconds)))
 
-
 @dataclass
-class TimedText:
+class Timed:
     start: Optional[float] = 0
     end: Optional[float] = 0
+
+@dataclass
+class TimedText(Timed):
     text: Optional[str] = ''
     speaker: Optional[int] = -1
-    probability: Optional[float] = None
-    is_dummy: Optional[bool] = False
     detected_language: Optional[str] = None
     
-    def is_punctuation(self):
-        return self.text.strip() in PUNCTUATION_MARKS
-    
-    def overlaps_with(self, other: 'TimedText') -> bool:
-        return not (self.end <= other.start or other.end <= self.start)
+    def has_punctuation(self) -> bool:
+        return any(char in PUNCTUATION_MARKS for char in self.text.strip())
     
     def is_within(self, other: 'TimedText') -> bool:
         return other.contains_timespan(self)
@@ -31,27 +28,26 @@ class TimedText:
     def duration(self) -> float:
         return self.end - self.start
 
-    def contains_time(self, time: float) -> bool:
-        return self.start <= time <= self.end
-
     def contains_timespan(self, other: 'TimedText') -> bool:
         return self.start <= other.start and self.end >= other.end
     
-    def __bool__(self):
+    def __bool__(self) -> bool:
         return bool(self.text)
-
+    
+    def __str__(self) -> str:
+        return str(self.text)
 
 @dataclass()
 class ASRToken(TimedText):
-    
-    corrected_speaker: Optional[int] = -1
-    validated_speaker: bool = False
-    validated_text: bool = False
-    validated_language: bool = False
-    
+    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, self.probability, 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
+
 
 @dataclass
 class Sentence(TimedText):
@@ -70,68 +66,93 @@ class Transcript(TimedText):
         sep: Optional[str] = None,
         offset: float = 0
     ) -> "Transcript":
+        """Collapse multiple ASR tokens into a single transcript span."""
         sep = sep if sep is not None else ' '
         text = sep.join(token.text for token in tokens)
-        probability = sum(token.probability for token in tokens if token.probability) / len(tokens) if tokens else None
         if tokens:
             start = offset + tokens[0].start
             end = offset + tokens[-1].end
         else:
             start = None
             end = None
-        return cls(start, end, text, probability=probability)
+        return cls(start, end, text)
 
 
 @dataclass
-class SpeakerSegment(TimedText):
+class SpeakerSegment(Timed):
     """Represents a segment of audio attributed to a specific speaker.
     No text nor probability is associated with this segment.
     """
+    speaker: Optional[int] = -1
     pass
 
 @dataclass
 class Translation(TimedText):
     pass
 
-    def approximate_cut_at(self, cut_time):
-        """
-        Each word in text is considered to be of duration (end-start)/len(words in text)
-        """
-        if not self.text or not self.contains_time(cut_time):
-            return self, None
-
-        words = self.text.split()
-        num_words = len(words)
-        if num_words == 0:
-            return self, None
-
-        duration_per_word = self.duration() / num_words
-        
-        cut_word_index = int((cut_time - self.start) / duration_per_word)
-        
-        if cut_word_index >= num_words:
-            cut_word_index = num_words -1
-        
-        text0 = " ".join(words[:cut_word_index])
-        text1 = " ".join(words[cut_word_index:])
-
-        segment0 = Translation(start=self.start, end=cut_time, text=text0)
-        segment1 = Translation(start=cut_time, end=self.end, text=text1)
-
-        return segment0, segment1
-        
-
 @dataclass
 class Silence():
-    duration: float
-    
+    start: Optional[float] = None
+    end: Optional[float] = None
+    duration: Optional[float] = None
+    is_starting: bool = False
+    has_ended: bool = False
+
+    def compute_duration(self) -> Optional[float]:
+        if self.start is None or self.end is None:
+            return None
+        self.duration = self.end - self.start
+        return self.duration
     
+    def is_silence(self) -> bool:
+        return True
+
+
 @dataclass
-class Line(TimedText):
-    translation: str = ''
-    
-    def to_dict(self):
-        _dict = {
+class Segment(TimedText):
+    """Generic contiguous span built from tokens or silence markers."""
+    start: Optional[float]
+    end: Optional[float]
+    text: Optional[str]
+    speaker: Optional[str]
+    tokens: Optional[ASRToken] = None
+    translation: Optional[Translation] = None
+
+    @classmethod
+    def from_tokens(
+        cls,
+        tokens: List[Union[ASRToken, Silence]],
+        is_silence: bool = False
+    ) -> Optional["Segment"]:
+        """Return a normalized segment representing the provided tokens."""
+        if not tokens:
+            return None
+        
+        start_token = tokens[0]
+        end_token = tokens[-1]        
+        if is_silence:
+            return cls(
+                start=start_token.start,
+                end=end_token.end,
+                text=None,
+                speaker=-2
+            )
+        else:
+            return cls(
+                start=start_token.start,
+                end=end_token.end,
+                text=''.join(token.text for token in tokens),
+                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
+
+    def to_dict(self) -> Dict[str, Any]:
+        """Serialize the segment for frontend consumption."""
+        _dict: Dict[str, Any] = {
             'speaker': int(self.speaker) if self.speaker != -1 else 1,
             'text': self.text,
             'start': format_time(self.start),
@@ -142,21 +163,33 @@ class Line(TimedText):
         if self.detected_language:
             _dict['detected_language'] = self.detected_language
         return _dict
-    
+
+
+@dataclass
+class PuncSegment(Segment):
+    pass
+
+class SilentSegment(Segment):
+    def __init__(self, *args: Any, **kwargs: Any) -> None:
+        super().__init__(*args, **kwargs)
+        self.speaker = -2
+        self.text = ''
+
 
 @dataclass  
 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 = ''
     remaining_time_transcription: float = 0.
     remaining_time_diarization: float = 0.
     
-    def to_dict(self):
-        _dict = {
+    def to_dict(self) -> Dict[str, Any]:
+        """Serialize the front-end data payload."""
+        _dict: Dict[str, Any] = {
             'status': self.status,
             'lines': [line.to_dict() for line in self.lines if (line.text or line.speaker == -2)],
             'buffer_transcription': self.buffer_transcription,
@@ -176,14 +209,22 @@ class ChangeSpeaker:
 
 @dataclass  
 class State():
-    tokens: list = field(default_factory=list)
-    last_validated_token: int = 0
-    last_punctuation_index: Optional[int] = None
-    translation_validated_segments: list = field(default_factory=list)
-    buffer_translation: str = field(default_factory=Transcript)
-    buffer_transcription: str = field(default_factory=Transcript)
+    """Unified state class for audio processing.
+    
+    Contains both persistent state (tokens, buffers) and temporary update buffers
+    (new_* fields) that are consumed by TokensAlignment.
+    """
+    # Persistent state
+    tokens: List[ASRToken] = field(default_factory=list)
+    buffer_transcription: Transcript = field(default_factory=Transcript)
     end_buffer: float = 0.0
     end_attributed_speaker: float = 0.0
     remaining_time_transcription: float = 0.0
     remaining_time_diarization: float = 0.0
-    beg_loop: Optional[int] = None
+    
+    # Temporary update buffers (consumed by TokensAlignment.update())
+    new_tokens: List[Union[ASRToken, Silence]] = field(default_factory=list)
+    new_translation: List[Any] = field(default_factory=list)
+    new_diarization: List[Any] = field(default_factory=list)
+    new_tokens_buffer: List[Any] = field(default_factory=list)  # only when local agreement
+    new_translation_buffer= TimedText()

whisperlivekit/tokens_alignment.py

@@ -0,0 +1,220 @@
+from time import time
+from typing import Any, List, Optional, Tuple, Union
+
+from whisperlivekit.timed_objects import (ASRToken, Segment, PuncSegment, Silence,
+                                          SilentSegment, SpeakerSegment,
+                                          TimedText)
+
+
+class TokensAlignment:
+
+    def __init__(self, state: Any, args: Any, sep: Optional[str]) -> None:
+        self.state = state
+        self.diarization = args.diarization
+        self._tokens_index: int = 0
+        self._diarization_index: int = 0
+        self._translation_index: int = 0
+
+        self.all_tokens: List[ASRToken] = []
+        self.all_diarization_segments: List[SpeakerSegment] = []
+        self.all_translation_segments: List[Any] = []
+
+        self.new_tokens: List[ASRToken] = []
+        self.new_diarization: List[SpeakerSegment] = []
+        self.new_translation: List[Any] = []
+        self.new_translation_buffer: Union[TimedText, str] = TimedText()
+        self.new_tokens_buffer: List[Any] = []
+        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, []
+        self.new_diarization, self.state.new_diarization = self.state.new_diarization, []
+        self.new_translation, self.state.new_translation = self.state.new_translation, []
+        self.new_tokens_buffer, self.state.new_tokens_buffer = self.state.new_tokens_buffer, []
+
+        self.all_tokens.extend(self.new_tokens)
+        self.all_diarization_segments.extend(self.new_diarization)
+        self.all_translation_segments.extend(self.new_translation)
+        self.new_translation_buffer = self.state.new_translation_buffer
+
+    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(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[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 = PuncSegment.from_tokens(
+                        tokens=self.all_tokens[segment_start_idx: i],
+                    )
+                if previous_segment:
+                    segments.append(previous_segment)
+                segment = PuncSegment.from_tokens(
+                    tokens=[token],
+                    is_silence=True
+                )
+                segments.append(segment)
+                segment_start_idx = i+1
+            else:
+                if token.has_punctuation():
+                    segment = PuncSegment.from_tokens(
+                        tokens=self.all_tokens[segment_start_idx: i+1],
+                    )
+                    segments.append(segment)
+                    segment_start_idx = i+1
+
+        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."""
+        if not self.all_diarization_segments:
+            return []
+        merged = [self.all_diarization_segments[0]]
+        for segment in self.all_diarization_segments[1:]:
+            if segment.speaker == merged[-1].speaker:
+                merged[-1].end = segment.end
+            else:
+                merged.append(segment)
+        return merged
+
+
+    @staticmethod
+    def intersection_duration(seg1: TimedText, seg2: TimedText) -> float:
+        """Return the overlap duration between two timed segments."""
+        start = max(seg1.start, seg2.start)
+        end = min(seg1.end, seg2.end)
+
+        return max(0, end - start)
+
+    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()
+        for punctuation_segment in punctuation_segments:
+            if not punctuation_segment.is_silence():
+                if diarization_segments and punctuation_segment.start >= diarization_segments[-1].end:
+                    diarization_buffer += punctuation_segment.text
+                else:
+                    max_overlap = 0.0
+                    max_overlap_speaker = 1
+                    for diarization_segment in diarization_segments:
+                        intersec = self.intersection_duration(punctuation_segment, diarization_segment)
+                        if intersec > max_overlap:
+                            max_overlap = intersec
+                            max_overlap_speaker = diarization_segment.speaker + 1
+                    punctuation_segment.speaker = max_overlap_speaker
+        
+        segments = []
+        if punctuation_segments:
+            segments = [punctuation_segments[0]]
+            for segment in punctuation_segments[1:]:
+                if segment.speaker == segments[-1].speaker:
+                    if segments[-1].text:
+                        segments[-1].text += segment.text
+                    segments[-1].end = segment.end
+                else:
+                    segments.append(segment)
+
+        return segments, diarization_buffer
+
+
+    def get_lines(
+            self, 
+            diarization: bool = False,
+            translation: bool = False,
+            current_silence: Optional[Silence] = None
+        ) -> Tuple[List[Segment], str, Union[str, TimedText]]:
+        """Return the formatted segments plus buffers, optionally with diarization/translation."""
+        if diarization:
+            segments, diarization_buffer = self.get_lines_diarization()
+        else:
+            diarization_buffer = ''
+            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 self.validated_segments and self.validated_segments[-1].is_silence():
+                        self.validated_segments[-1].end = end_silence
+                    else:
+                        self.validated_segments.append(SilentSegment(
+                            start=token.start,
+                            end=end_silence
+                        ))
+                else:
+                    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 segments and segments[-1].is_silence():
+                segments[-1] = SilentSegment(start=segments[-1].start, end=end_silence)
+            else:
+                segments.append(SilentSegment(
+                    start=current_silence.start,
+                    end=end_silence
+                ))
+        if translation:
+            [self.add_translation(segment) for segment in segments if not segment.is_silence()]
+        return segments, diarization_buffer, self.new_translation_buffer.text

whisperlivekit/trail_repetition.py

@@ -1,60 +0,0 @@
-from typing import Sequence, Callable, Any, Optional, Dict
-
-def _detect_tail_repetition(
-    seq: Sequence[Any],
-    key: Callable[[Any], Any] = lambda x: x,  # extract comparable value
-    min_block: int = 1,                       # set to 2 to ignore 1-token loops like "."
-    max_tail: int = 300,                      # search window from the end for speed
-    prefer: str = "longest",                  # "longest" coverage or "smallest" block
-) -> Optional[Dict]:
-    vals = [key(x) for x in seq][-max_tail:]
-    n = len(vals)
-    best = None
-
-    # try every possible block length
-    for b in range(min_block, n // 2 + 1):
-        block = vals[-b:]
-        # count how many times this block repeats contiguously at the very end
-        count, i = 0, n
-        while i - b >= 0 and vals[i - b:i] == block:
-            count += 1
-            i -= b
-
-        if count >= 2:
-            cand = {
-                "block_size": b,
-                "count": count,
-                "start_index": len(seq) - count * b,  # in original seq
-                "end_index": len(seq),
-            }
-            if (best is None or
-                (prefer == "longest" and count * b > best["count"] * best["block_size"]) or
-                (prefer == "smallest" and b < best["block_size"])):
-                best = cand
-    return best
-
-def trim_tail_repetition(
-    seq: Sequence[Any],
-    key: Callable[[Any], Any] = lambda x: x,
-    min_block: int = 1,
-    max_tail: int = 300,
-    prefer: str = "longest",
-    keep: int = 1,  # how many copies of the repeating block to keep at the end (0 or 1 are common)
-):
-    """
-    Returns a new sequence with repeated tail trimmed.
-    keep=1 -> keep a single copy of the repeated block.
-    keep=0 -> remove all copies of the repeated block.
-    """
-    rep = _detect_tail_repetition(seq, key, min_block, max_tail, prefer)
-    if not rep:
-        return seq, False  # nothing to trim
-
-    b, c = rep["block_size"], rep["count"]
-    if keep < 0:
-        keep = 0
-    if keep >= c:
-        return seq, False  # nothing to trim (already <= keep copies)
-    # new length = total - (copies_to_remove * block_size)
-    new_len = len(seq) - (c - keep) * b
-    return seq[:new_len], True

whisperlivekit/voxtral_streaming.py

@@ -0,0 +1,484 @@
+"""
+Voxtral Mini Realtime streaming backend using voxmlx's incremental encode/decode.
+
+Uses model.encode_step() for incremental audio encoding and token-by-token
+autoregressive decoding, matching voxmlx's native streaming pipeline.
+"""
+
+import logging
+import sys
+import time
+from typing import List, Optional, Tuple
+
+import numpy as np
+
+from whisperlivekit.timed_objects import ASRToken, Transcript
+
+logger = logging.getLogger(__name__)
+
+N_LEFT_PAD_TOKENS = 32
+N_RIGHT_PAD_TOKENS = 17
+
+
+class VoxtralStreamingASR:
+    """Voxtral model holder for the streaming pipeline."""
+
+    sep = " "
+
+    def __init__(self, logfile=sys.stderr, **kwargs):
+        from voxmlx import _build_prompt_tokens
+        from voxmlx import load_model as vox_load_model
+
+        self.logfile = logfile
+        self.transcribe_kargs = {}
+
+        lan = kwargs.get("lan", "auto")
+        self.original_language = None if lan == "auto" else lan
+
+        DEFAULT_MODEL = "mlx-community/Voxtral-Mini-4B-Realtime-6bit"
+        model_path = kwargs.get("model_dir") or kwargs.get("model_path")
+        if not model_path:
+            model_size = kwargs.get("model_size", "")
+            # Only use model_size if it looks like a HF repo or a path, not a Whisper size name
+            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 voxmlx...")
+        self.model, self._tokenizer, self._config = vox_load_model(model_path)
+        self._prompt_tokens, self._n_delay_tokens = _build_prompt_tokens(
+            self._tokenizer
+        )
+        logger.info(f"Voxtral model loaded in {time.time() - t:.2f}s")
+
+        self.backend_choice = "voxtral-mlx"
+        self.tokenizer = None  # sentence tokenizer — not needed for streaming
+
+    def transcribe(self, audio):
+        pass
+
+
+class VoxtralStreamingOnlineProcessor:
+    """
+    Online processor for Voxtral streaming ASR.
+
+    Uses voxmlx's incremental encoding (encode_step) and token-by-token
+    autoregressive decoding. Each decode step corresponds to 80ms of audio.
+    """
+
+    SAMPLING_RATE = 16000
+
+    def __init__(self, asr: VoxtralStreamingASR, logfile=sys.stderr):
+        from mistral_common.tokens.tokenizers.base import SpecialTokenPolicy
+
+        self.asr = asr
+        self.logfile = logfile
+        self.end = 0.0
+        self.buffer = []
+        self.audio_buffer = np.array([], dtype=np.float32)  # for logging compat
+        self._special_token_policy = SpecialTokenPolicy.IGNORE
+        self._reset_state()
+        logger.info(
+            f"[voxtral] Initialized. eos_id={asr._tokenizer.eos_id}, "
+            f"prefix_len={len(asr._prompt_tokens)}, "
+            f"n_delay={asr._n_delay_tokens}"
+        )
+
+    def _reset_state(self):
+        from voxmlx.audio import SAMPLES_PER_TOKEN
+
+        self._samples_per_token = SAMPLES_PER_TOKEN
+
+        # Incremental encoder state
+        self._audio_tail = None
+        self._conv1_tail = None
+        self._conv2_tail = None
+        self._encoder_cache = None
+        self._ds_buf = None
+
+        # Decoder state
+        self._decoder_cache = None
+        self._y = None  # last sampled token (mx.array scalar)
+        self._t_cond = None
+        self._text_embeds = None
+
+        # Audio / decode tracking
+        self._pending_audio = np.zeros(0, dtype=np.float32)
+        self._audio_embeds = None
+        self._n_audio_samples_fed = 0
+        self._n_total_decoded = 0
+        self._first_cycle = True
+        self._prefilled = False
+
+        # Word extraction: accumulate token IDs, full-sequence decode for correct spacing
+        self._output_token_ids: List[int] = []
+        self._token_positions: List[int] = []  # decode position for each token
+        self._n_committed_words = 0
+        self._global_time_offset = 0.0
+        self._y_flushed_to_output = False  # True after start_silence flushes pending _y
+
+    # ── Interface methods (same as SimulStreamingOnlineProcessor) ──
+
+    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  # for logging compat
+
+    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] process_iter exception: {e}", exc_info=True)
+            return [], self.end
+
+    def _get_full_text(self) -> str:
+        """Decode all accumulated token IDs at once for correct spacing."""
+        if not self._output_token_ids:
+            return ""
+        sp = self.asr._tokenizer
+        return sp.decode(self._output_token_ids, special_token_policy=self._special_token_policy)
+
+    def get_buffer(self) -> Transcript:
+        """Return all uncommitted text as buffer, including pending _y token."""
+        # Temporarily include pending _y for buffer display
+        ids = list(self._output_token_ids)
+        if self._y is not None and not self._y_flushed_to_output:
+            sp = self.asr._tokenizer
+            token_id = self._y.item()
+            if token_id != sp.eos_id:
+                ids.append(token_id)
+        if not ids:
+            return Transcript(start=None, end=None, text="")
+        sp = self.asr._tokenizer
+        full_text = sp.decode(ids, special_token_policy=self._special_token_policy)
+        words = full_text.split()
+        uncommitted = words[self._n_committed_words:]
+        if uncommitted:
+            text = " ".join(uncommitted)
+            return Transcript(start=self.end, end=self.end, text=text)
+        return Transcript(start=None, end=None, text="")
+
+    def start_silence(self) -> Tuple[List[ASRToken], float]:
+        """Flush all uncommitted words when silence starts."""
+        self._flush_last_y()  # Include the pending _y token before flushing
+        words = self._flush_all_pending_words()
+        logger.info(f"[voxtral] 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 to let the model finish decoding."""
+        right_pad = np.zeros(
+            N_RIGHT_PAD_TOKENS * self._samples_per_token, dtype=np.float32
+        )
+        self._pending_audio = np.append(self._pending_audio, right_pad)
+        self._n_audio_samples_fed += len(right_pad)
+
+        final_words, _ = self._process_iter_inner(is_last=True)
+        # Flush the last pending self._y token (like voxmlx's finally block)
+        self._flush_last_y()
+        final_words.extend(self._flush_all_pending_words())
+        return final_words, self.end
+
+    # ── Word extraction ──
+
+    def _pos_to_time(self, pos: int) -> float:
+        """Convert a decode position to seconds relative to audio start."""
+        SPT = self._samples_per_token
+        return max(0.0, (pos - N_LEFT_PAD_TOKENS) * SPT / self.SAMPLING_RATE)
+
+    def _flush_last_y(self):
+        """Flush the last pending self._y token that hasn't been processed yet."""
+        if self._y is None or self._y_flushed_to_output:
+            return
+        sp = self.asr._tokenizer
+        token_id = self._y.item()
+        if token_id != sp.eos_id:
+            self._output_token_ids.append(token_id)
+            self._token_positions.append(self._n_total_decoded)
+            self._y_flushed_to_output = True
+
+    def _extract_new_words(self) -> List[ASRToken]:
+        """
+        Split accumulated text into words and return new complete words
+        (all but the last, which may still be growing).
+        """
+        if not self._output_token_ids:
+            return []
+
+        full_text = self._get_full_text()
+        words = full_text.split()
+
+        new_words: List[ASRToken] = []
+        n_tokens = len(self._output_token_ids)
+        # All words except the last are guaranteed complete
+        while len(words) > self._n_committed_words + 1:
+            word = words[self._n_committed_words]
+            word_idx = self._n_committed_words
+            n_words_total = len(words)
+            # Approximate: assign token range proportionally
+            tok_start = int(word_idx / n_words_total * n_tokens)
+            tok_end = int((word_idx + 1) / n_words_total * n_tokens)
+            tok_start = min(tok_start, len(self._token_positions) - 1)
+            tok_end = min(tok_end, len(self._token_positions) - 1)
+
+            start_time = self._pos_to_time(self._token_positions[tok_start]) + self._global_time_offset
+            end_time = self._pos_to_time(self._token_positions[tok_end]) + self._global_time_offset
+
+            # Prepend space to match Whisper convention (Segment.from_tokens joins with '')
+            text = word if self._n_committed_words == 0 else " " + word
+            new_words.append(ASRToken(start=start_time, end=end_time, text=text))
+            self._n_committed_words += 1
+
+        return new_words
+
+    def _flush_all_pending_words(self) -> List[ASRToken]:
+        """Flush ALL words including the last partial one."""
+        if not self._output_token_ids:
+            return []
+
+        full_text = self._get_full_text()
+        words = full_text.split()
+
+        new_words: List[ASRToken] = []
+        n_tokens = len(self._output_token_ids)
+        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)
+            tok_start = min(tok_start, max(len(self._token_positions) - 1, 0))
+            tok_end = min(tok_end, max(len(self._token_positions) - 1, 0))
+
+            if self._token_positions:
+                start_time = self._pos_to_time(self._token_positions[tok_start]) + self._global_time_offset
+                end_time = self._pos_to_time(self._token_positions[tok_end]) + self._global_time_offset
+            else:
+                start_time = self._global_time_offset
+                end_time = self._global_time_offset
+
+            # Prepend space to match Whisper convention (Segment.from_tokens joins with '')
+            text = word if self._n_committed_words == 0 else " " + word
+            new_words.append(ASRToken(start=start_time, end=end_time, text=text))
+            self._n_committed_words += 1
+
+        return new_words
+
+    # ── Core streaming logic ──
+
+    def _process_iter_inner(self, is_last: bool) -> Tuple[List[ASRToken], float]:
+        import mlx.core as mx
+
+        from voxmlx.audio import log_mel_spectrogram_step
+        from voxmlx.cache import RotatingKVCache
+
+        model = self.asr.model
+        sp = self.asr._tokenizer
+        prompt_tokens = self.asr._prompt_tokens
+        prefix_len = len(prompt_tokens)
+        SPT = self._samples_per_token
+
+        # ── Phase 1: Encode new audio ──
+        if self._first_cycle and len(self._pending_audio) >= SPT:
+            left_pad = np.zeros(N_LEFT_PAD_TOKENS * SPT, dtype=np.float32)
+            n_feed = (len(self._pending_audio) // SPT) * SPT
+            chunk = np.concatenate([left_pad, self._pending_audio[:n_feed]])
+            self._pending_audio = self._pending_audio[n_feed:]
+            self._n_audio_samples_fed += n_feed
+
+            mel, self._audio_tail = log_mel_spectrogram_step(
+                chunk, self._audio_tail
+            )
+            (
+                new_embeds,
+                self._conv1_tail,
+                self._conv2_tail,
+                self._encoder_cache,
+                self._ds_buf,
+            ) = model.encode_step(
+                mel,
+                self._conv1_tail,
+                self._conv2_tail,
+                self._encoder_cache,
+                self._ds_buf,
+            )
+            if new_embeds is not None:
+                mx.eval(new_embeds)
+                self._audio_embeds = new_embeds
+                logger.info(f"[voxtral] first encode: {new_embeds.shape[0]} embeds from {n_feed} samples")
+            else:
+                logger.info(f"[voxtral] first encode: no embeds from {n_feed} samples")
+            self._first_cycle = False
+
+        elif not self._first_cycle and len(self._pending_audio) >= SPT:
+            n_feed = (len(self._pending_audio) // SPT) * SPT
+            chunk = self._pending_audio[:n_feed]
+            self._pending_audio = self._pending_audio[n_feed:]
+            self._n_audio_samples_fed += n_feed
+
+            mel, self._audio_tail = log_mel_spectrogram_step(
+                chunk, self._audio_tail
+            )
+            (
+                new_embeds,
+                self._conv1_tail,
+                self._conv2_tail,
+                self._encoder_cache,
+                self._ds_buf,
+            ) = model.encode_step(
+                mel,
+                self._conv1_tail,
+                self._conv2_tail,
+                self._encoder_cache,
+                self._ds_buf,
+            )
+            if new_embeds is not None:
+                mx.eval(new_embeds)
+                if self._audio_embeds is not None:
+                    self._audio_embeds = mx.concatenate(
+                        [self._audio_embeds, new_embeds]
+                    )
+                else:
+                    self._audio_embeds = new_embeds
+
+        self.audio_buffer = self._pending_audio  # for logging compat
+
+        if self._audio_embeds is None:
+            return [], self.end
+
+        # Safety: don't decode ahead of encoded audio
+        safe_total = (
+            N_LEFT_PAD_TOKENS + self._n_audio_samples_fed // SPT
+        )
+        n_decodable = min(
+            self._audio_embeds.shape[0], safe_total - self._n_total_decoded
+        )
+
+        if n_decodable <= 0:
+            return [], self.end
+
+        # ── Phase 2: Prefill (once per utterance) ──
+        if not self._prefilled:
+            if self._n_total_decoded + self._audio_embeds.shape[0] < prefix_len:
+                logger.info(
+                    f"[voxtral] waiting for prefill: have {self._audio_embeds.shape[0]} embeds, need {prefix_len}"
+                )
+                return [], self.end
+
+            n_layers = len(model.language_model.layers)
+            self._decoder_cache = [RotatingKVCache(8192) for _ in range(n_layers)]
+
+            self._t_cond = model.time_embedding(
+                mx.array([self.asr._n_delay_tokens], dtype=mx.float32)
+            )
+
+            prompt_ids = mx.array([prompt_tokens])
+            self._text_embeds = model.language_model.embed(prompt_ids)[0]
+
+            prefix_embeds = (
+                self._text_embeds + self._audio_embeds[:prefix_len]
+            )[None, :, :]
+
+            logits = model.decode(
+                prefix_embeds, self._t_cond, "causal", self._decoder_cache
+            )
+            mx.eval(
+                logits,
+                *[x for c in self._decoder_cache for x in (c.keys, c.values)],
+            )
+
+            self._y = mx.argmax(logits[0, -1:], axis=-1).squeeze()
+            mx.async_eval(self._y)
+
+            self._audio_embeds = self._audio_embeds[prefix_len:]
+            self._n_total_decoded = prefix_len
+            self._prefilled = True
+            logger.info(f"[voxtral] prefill done, first token y={self._y.item()}")
+
+            n_decodable = min(
+                self._audio_embeds.shape[0], safe_total - self._n_total_decoded
+            )
+
+        if n_decodable <= 0:
+            return [], self.end
+
+        # ── Phase 3: Decode new positions ──
+        eos_id = sp.eos_id
+        hit_eos = False
+        n_consumed = 0
+
+        for i in range(n_decodable):
+            token_embed = model.language_model.embed(self._y.reshape(1, 1))[0, 0]
+            step_embed = (self._audio_embeds[i] + token_embed)[None, None, :]
+            logits = model.decode(
+                step_embed, self._t_cond, mask=None, cache=self._decoder_cache
+            )
+            next_y = mx.argmax(logits[0, -1:], axis=-1).squeeze()
+            mx.async_eval(next_y)
+
+            token_id = self._y.item()
+            n_consumed = i + 1
+
+            if token_id == eos_id:
+                hit_eos = True
+                logger.info("[voxtral] hit EOS")
+                break
+
+            # Accumulate token ID — full-sequence decode produces correct spacing
+            # Skip if this _y was already flushed by start_silence()
+            if self._y_flushed_to_output:
+                self._y_flushed_to_output = False
+            else:
+                self._output_token_ids.append(token_id)
+                # Track position for timestamp estimation
+                pos = self._n_total_decoded + i
+                self._token_positions.append(pos)
+
+            if i > 0 and i % 256 == 0:
+                mx.clear_cache()
+
+            self._y = next_y
+
+        self._n_total_decoded += n_consumed
+
+        # Trim consumed embeddings
+        if self._audio_embeds.shape[0] > n_consumed:
+            self._audio_embeds = self._audio_embeds[n_consumed:]
+        else:
+            self._audio_embeds = None
+
+        # Log decode results
+        full_text = self._get_full_text()
+        logger.info(
+            f"[voxtral] decoded {n_consumed} tokens | "
+            f"total_decoded={self._n_total_decoded} | "
+            f"text='{full_text[-80:]}' | "
+            f"n_words={len(full_text.split())} committed={self._n_committed_words}"
+        )
+
+        # Extract complete words from the decoded token sequence
+        new_words = self._extract_new_words()
+
+        if hit_eos:
+            new_words.extend(self._flush_all_pending_words())
+            self._reset_state()
+
+        if new_words:
+            logger.info(f"[voxtral] returning {len(new_words)} words: {[w.text for w in new_words]}")
+
+        self.buffer = []
+        return new_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/live_transcription.js

@@ -391,12 +391,11 @@ function renderLinesWithBuffer(
 
       if (idx === lines.length - 1) {
         if (!isFinalizing && item.speaker !== -2) {
-          if (remaining_time_transcription > 0) {
             speakerLabel += `<span class="label_transcription"><span class="spinner"></span>Transcription lag <span id='timeInfo'><span class="lag-transcription-value">${fmt1(
               remaining_time_transcription
             )}</span>s</span></span>`;
-          }
-          if (buffer_diarization && remaining_time_diarization > 0) {
+
+          if (buffer_diarization && remaining_time_diarization) {
             speakerLabel += `<span class="label_diarization"><span class="spinner"></span>Diarization lag<span id='timeInfo'><span class="lag-diarization-value">${fmt1(
               remaining_time_diarization
             )}</span>s</span></span>`;

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

@@ -4,17 +4,20 @@ import json
 import os
 import urllib
 import warnings
-from typing import List, Optional, Union, Dict
+from pathlib import Path
+from typing import Dict, List, Optional, Union
 
 import torch
+from torch import Tensor
 from tqdm import tqdm
-from pathlib import Path
 
-from .audio import load_audio, log_mel_spectrogram, pad_or_trim
-from .decoding import DecodingOptions, DecodingResult, decode, detect_language
-from .model import ModelDimensions, Whisper
-from .transcribe import transcribe
-from .version import __version__
+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__
 
 _MODELS = {
     "tiny.en": "https://openaipublic.azureedge.net/main/whisper/models/d3dd57d32accea0b295c96e26691aa14d8822fac7d9d27d5dc00b4ca2826dd03/tiny.en.pt",
@@ -105,7 +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):
@@ -119,6 +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"],
@@ -233,13 +255,224 @@ 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")
+    if os.path.isfile(safe_path):
+        try:
+            from safetensors.torch import load_file
+        except ImportError as exc:
+            raise ImportError(
+                "Loading LoRA adapters stored as .safetensors requires the `safetensors` package."
+            ) from exc
+        return load_file(safe_path)
+    if os.path.isfile(bin_path):
+        return torch.load(bin_path, map_location="cpu")
+    raise FileNotFoundError(
+        f"No adapter weights found under {lora_path}. Expected adapter_model.safetensors or adapter_model.bin."
+    )
+
+
+def _collapse_hf_module_name(module: str):
+    if module.startswith("base_model."):
+        module = module[len("base_model.") :]
+    if module.startswith("model.model."):
+        module = module[len("model.") :]
+    if not module.startswith("model."):
+        module = f"model.{module}"
+    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):
+        raise FileNotFoundError(f"Missing adapter_config.json inside {lora_path}")
+    with open(config_path, "r", encoding="utf-8") as handle:
+        config = json.load(handle)
+    if config.get("peft_type") != "LORA":
+        raise ValueError("Only LoRA adapters are supported.")
+
+    r = config.get("r")
+    alpha = config.get("lora_alpha") or config.get("alpha")
+    if not r or not alpha:
+        raise ValueError("LoRA config must include `r` and `lora_alpha`.")
+    scaling = alpha / r
+
+    adapter_state = _load_lora_state(lora_path)
+    lora_layers: Dict[str, Dict[str, Tensor]] = {}
+    for key, tensor in adapter_state.items():
+        if key.endswith("lora_A.weight"):
+            module = key[: -len(".lora_A.weight")]
+            lora_layers.setdefault(module, {})["A"] = tensor
+        elif key.endswith("lora_B.weight"):
+            module = key[: -len(".lora_B.weight")]
+            lora_layers.setdefault(module, {})["B"] = tensor
+
+    if not lora_layers:
+        raise ValueError(f"No LoRA tensors found in {lora_path}")
+
+    for module, parts in lora_layers.items():
+        if "A" not in parts or "B" not in parts:
+            raise ValueError(f"Incomplete LoRA tensors for module '{module}'")
+
+        hf_module = _collapse_hf_module_name(module)
+        hf_weight_key = f"{hf_module}.weight"
+
+        delta = parts["B"] @ parts["A"]
+        delta = delta * scaling
+
+        converted = _convert_hf_state_dict({hf_weight_key: delta})
+        if not converted:
+            raise KeyError(f"Failed to map LoRA module '{module}' into Whisper state dict.")
+        target_name, delta_tensor = next(iter(converted.items()))
+        if target_name not in state_dict:
+            raise KeyError(
+                f"LoRA module '{module}' mapped to '{target_name}', but the base model has no such parameter."
+            )
+
+        state_dict[target_name] = state_dict[target_name] + delta_tensor.to(
+            dtype=state_dict[target_name].dtype, device=state_dict[target_name].device
+        )
+
+
+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,
     download_root: str = None,
     in_memory: bool = False,
-    decoder_only=False,
-    custom_alignment_heads=None
+    decoder_only: bool = False,
+    custom_alignment_heads: Optional[str] = None,
+    lora_path: Optional[str] = None,
 ) -> Whisper:
     """
     Load a Whisper ASR model
@@ -249,28 +482,67 @@ 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
         path to download the model files; by default, it uses "~/.cache/whisper"
     in_memory: bool
         whether to preload the model weights into host memory
+    lora_path: str
+        optional directory containing PEFT LoRA adapter weights (adapter_config + adapter_model)
 
     Returns
     -------
     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()}"
@@ -280,33 +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. "
@@ -326,6 +588,55 @@ 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)
+
+
+def convert_encoder_to_coreml(
+    model_name = "base",
+    output_path= "whisper_encoder.mlpackage",
+    dummy_frames = 3000, #Number of time frames to use for the dummy mel input during tracing
+    precision = "float16",
+):
+   
+    import coremltools as ct
+    model = load_model(model_name, device="cpu", decoder_only=False)
+    encoder = model.encoder.eval().cpu()
+
+    dummy_input = torch.randn(
+        1,
+        model.dims.n_mels,
+        dummy_frames,
+        dtype=next(encoder.parameters()).dtype,
+    )
+
+    with torch.no_grad():
+        traced_encoder = torch.jit.trace(encoder, dummy_input)
+
+    precision_map = {
+        "float16": ct.precision.FLOAT16,
+        "fp16": ct.precision.FLOAT16,
+        "float32": ct.precision.FLOAT32,
+        "fp32": ct.precision.FLOAT32,
+    }
+    coreml_precision = precision_map[precision.lower()]
+
+    mlmodel = ct.convert(
+        traced_encoder,
+        inputs=[ct.TensorType(name="mel", shape=dummy_input.shape)],
+        convert_to= "mlprogram",
+        compute_precision=coreml_precision,
+    )
+
+    output_path = Path(output_path)
+    mlmodel.save(str(output_path))
+    return output_path
+
+# if __name__ == "__main__":
+#     convert_encoder_to_coreml(model_name="tiny", output_path="whisper_encoder.mlpackage", dummy_frames=3000, precision="float16", convert_to="mlprogram")

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/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