0.2.6

frontend: Keep screen awake and scroll down when transcribing.

.gitignore

@@ -54,7 +54,6 @@ coverage.xml
 # Translations
 *.mo
 *.pot
-
 # Django stuff:
 *.log
 local_settings.py
@@ -129,4 +128,13 @@ dmypy.json
 .pyre/
 
 *.wav
-run_*.sh
+run_*.sh
+
+# Downloaded models
+*.pt 
+
+# Debug & testing
+test_*.py
+launch.json
+.DS_Store
+test/*

CONTRIBUTING.md

@@ -15,7 +15,7 @@ Thank you for considering contributing ! We appreciate your time and effort to h
 
 ## Opening Issues
 
-If you encounter a problem with diart or want to suggest an improvement, please follow these guidelines when opening an issue:
+If you encounter a problem with WhisperLiveKit or want to suggest an improvement, please follow these guidelines when opening an issue:
 
 - **Bug Reports:**
   - Clearly describe the error. **Please indicate the parameters you use, especially the model(s)**
@@ -43,4 +43,4 @@ We welcome and appreciate contributions! To ensure a smooth review process, plea
 
 ## Thank You
 
-Your contributions make diart better for everyone. Thank you for your time and dedication!
+Your contributions make WhisperLiveKit better for everyone. Thank you for your time and dedication!

Dockerfile

@@ -21,15 +21,17 @@ RUN apt-get update && \
         python3 \
         python3-pip \
         ffmpeg \
-        git && \
+        git \
+        build-essential \
+        python3-dev && \
     rm -rf /var/lib/apt/lists/*
 
-RUN pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
+RUN pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
 
 COPY . .
 
 # Install WhisperLiveKit directly, allowing for optional dependencies
-#   Note: For gates modedls, need to add your HF toke. See README.md
+#   Note: For gates models, need to add your HF toke. See README.md
 #         for more details.
 RUN if [ -n "$EXTRAS" ]; then \
       echo "Installing with extras: [$EXTRAS]"; \
@@ -79,4 +81,4 @@ EXPOSE 8000
 ENTRYPOINT ["whisperlivekit-server", "--host", "0.0.0.0"]
 
 # Default args
-CMD ["--model", "tiny.en"]
+CMD ["--model", "base"]

LICENSE

@@ -1,3 +1,7 @@
+# License
+
+## Main Software License
+
 MIT License
 
 Copyright (c) 2025 Quentin Fuxa.  
@@ -20,9 +24,29 @@ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE  
 SOFTWARE.
 
+## SimulStreaming Backend License
+
+**When using the SimulStreaming backend (SimulWhisper), additional licensing terms apply:**
+
+SimulStreaming (https://github.com/ufal/SimulStreaming) is dual-licensed:
+
+### 🔹 Non-Commercial Use
+You may use SimulStreaming under the **PolyForm Noncommercial License 1.0.0** if you obtain the code through the GitHub repository. This license is **free of charge** and comes with **no obligations** for non-commercial users.
+
+### 🔸 Commercial Use
+Understanding who uses SimulStreaming commercially helps improve and prioritize development. Therefore, **registration is required** for those who acquire a commercial license.
+
+Commercial licenses are planned to be **affordable** to SMEs and individuals. They are considering providing commercial licenses either for free or for a symbolic one-time fee, and may also provide additional support. You can share your preference via the [questionnaire](https://forms.cloud.microsoft.com/e/7tCxb4gJfB).
+
+You can also leave your contact [there](https://forms.cloud.microsoft.com/e/7tCxb4gJfB) to be notified when commercial licenses become available.
+
+**Contact for SimulStreaming licensing:**  
+[Dominik Macháček](https://ufal.mff.cuni.cz/dominik-machacek/), machacek@ufal.mff.cuni.cz
+
 ---
 
-Based on:
+## Based on:
 - **whisper_streaming** by ÚFAL – MIT License – https://github.com/ufal/whisper_streaming. The original work by ÚFAL. License: https://github.com/ufal/whisper_streaming/blob/main/LICENSE  
 - **silero-vad** by Snakers4 – MIT License – https://github.com/snakers4/silero-vad. The work by Snakers4 (silero-vad). License: https://github.com/snakers4/silero-vad/blob/f6b1294cb27590fb2452899df98fb234dfef1134/LICENSE  
-- **Diart** by juanmc2005 – MIT License – https://github.com/juanmc2005/diart. The work in Diart by juanmc2005. License: https://github.com/juanmc2005/diart/blob/main/LICENSE
+- **Diart** by juanmc2005 – MIT License – https://github.com/juanmc2005/diart. The work in Diart by juanmc2005. License: https://github.com/juanmc2005/diart/blob/main/LICENSE
+- **SimulStreaming** by ÚFAL – Dual License (PolyForm Noncommercial License 1.0.0 / Commercial License) – https://github.com/ufal/SimulStreaming

README.md

@@ -4,176 +4,125 @@
 <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 Diarization</b></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=downloads"></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-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>
 
-## 🚀 Overview
 
-This project is based on [WhisperStreaming](https://github.com/ufal/whisper_streaming) and [SimulStreaming](https://github.com/ufal/SimulStreaming), allowing you to transcribe audio directly from your browser. WhisperLiveKit provides a complete backend solution for real-time speech transcription with a functional, simple and customizable frontend. Everything runs locally on your machine ✨
+Real-time speech transcription directly to your browser, with a ready-to-use backend+server and a simple frontend. ✨
 
-### 🔄 Architecture
+#### Powered by Leading Research:
 
-WhisperLiveKit consists of three main components:
-
-- **Frontend**: A basic html + JS interface that captures microphone audio and streams it to the backend via WebSockets. You can use and adapt the provided template at [whisperlivekit/web/live_transcription.html](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/web/live_transcription.html).
-- **Backend (Web Server)**: A FastAPI-based WebSocket server that receives streamed audio data, processes it in real time, and returns transcriptions to the frontend. This is where the WebSocket logic and routing live.
-- **Core Backend (Library Logic)**: A server-agnostic core that handles audio processing, ASR, and diarization. It exposes reusable components that take in audio bytes and return transcriptions.
+- [SimulStreaming](https://github.com/ufal/SimulStreaming) (SOTA 2025) - Ultra-low latency transcription with AlignAtt policy
+- [WhisperStreaming](https://github.com/ufal/whisper_streaming) (SOTA 2023) - Low latency transcription with LocalAgreement policy
+- [Streaming Sortformer](https://arxiv.org/abs/2507.18446) (SOTA 2025) - Advanced real-time speaker diarization
+- [Diart](https://github.com/juanmc2005/diart) (SOTA 2021) - Real-time speaker diarization
+- [Silero VAD](https://github.com/snakers4/silero-vad) (2024) - Enterprise-grade Voice Activity Detection
 
 
-### ✨ Key Features
-
-- **🎙️ Real-time Transcription** - Locally (or on-prem) convert speech to text instantly as you speak
-- **👥 Speaker Diarization** - Identify different speakers in real-time using [Diart](https://github.com/juanmc2005/diart)
-- **🌐 Multi-User Support** - Handle multiple users simultaneously with a single backend/server
-- **🔇 Automatic Silence Chunking** – Automatically chunks when no audio is detected to limit buffer size
-- **✅ Confidence Validation** – Immediately validate high-confidence tokens for faster inference (WhisperStreaming only)
-- **👁️ Buffering Preview** – Displays unvalidated transcription segments (not compatible with SimulStreaming yet)
-- **✒️ Punctuation-Based Speaker Splitting [BETA]** - Align speaker changes with natural sentence boundaries for more readable transcripts
-- **⚡ SimulStreaming Backend** - Ultra-low latency transcription using state-of-the-art AlignAtt policy. The code is not directly included in the repo : To use, please copy [simul_whisper](https://github.com/ufal/SimulStreaming/tree/main/simul_whisper) content into `whisperlivekit/simul_whisper` . ⚠️ You must comply with the [Polyform license](https://github.com/ufal/SimulStreaming/blob/main/LICENCE.txt)
+> **Why not just run a simple Whisper model on every audio batch?** Whisper is designed for complete utterances, not real-time chunks. Processing small segments loses context, cuts off words mid-syllable, and produces poor transcription. WhisperLiveKit uses state-of-the-art simultaneous speech research for intelligent buffering and incremental processing.
 
 
-## 📖 Quick Start
+### Architecture
 
-```bash
-# Install the package
-pip install whisperlivekit
+<img alt="Architecture" src="https://raw.githubusercontent.com/QuentinFuxa/WhisperLiveKit/refs/heads/main/architecture.png" />
 
-# Start the transcription server
-whisperlivekit-server --model tiny.en
+*The backend supports multiple concurrent users. Voice Activity Detection reduces overhead when no voice is detected.*
 
-# Open your browser at http://localhost:8000 to see the interface.
-# Use  -ssl-certfile public.crt --ssl-keyfile private.key parameters to use SSL
-```
-
-That's it! Start speaking and watch your words appear on screen.
-
-## 🛠️ Installation Options
-
-### Install from PyPI (Recommended)
+### Installation & Quick Start
 
 ```bash
 pip install whisperlivekit
 ```
 
-### Install from Source
+>  **FFmpeg is required** and must be installed before using WhisperLiveKit
+> 
+> | OS | How to install |
+> |-----------|-------------|
+>  | Ubuntu/Debian | `sudo apt install ffmpeg` |
+> | MacOS | `brew install ffmpeg` |
+> | Windows | Download .exe from https://ffmpeg.org/download.html and add to PATH |
 
-```bash
-git clone https://github.com/QuentinFuxa/WhisperLiveKit
-cd WhisperLiveKit
-pip install -e .
-```
+#### Quick Start
+1. **Start the transcription server:**
+   ```bash
+   whisperlivekit-server --model base --language en
+   ```
 
-### System Dependencies
+2. **Open your browser** and navigate to `http://localhost:8000`. Start speaking and watch your words appear in real-time!
 
-FFmpeg is required:
 
-```bash
-# Ubuntu/Debian
-sudo apt install ffmpeg
+> - 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.
 
-# macOS
-brew install ffmpeg
+ 
 
-# Windows
-# Download from https://ffmpeg.org/download.html and add to PATH
-```
+#### Optional Dependencies
 
-### Optional Dependencies
+| Optional | `pip install` |
+|-----------|-------------|
+| Speaker diarization | `whisperlivekit[diarization]` |
+| Original Whisper backend | `whisperlivekit[whisper]` |
+| Improved timestamps backend | `whisperlivekit[whisper-timestamped]` |
+| Apple Silicon optimization backend | `whisperlivekit[mlx-whisper]` |
+| OpenAI API backend | `whisperlivekit[openai]` |
 
-```bash
-# Voice Activity Controller (prevents hallucinations)
-pip install torch
+See  **Parameters & Configuration** below on how to use them.
 
-# Sentence-based buffer trimming
-pip install mosestokenizer wtpsplit
-pip install tokenize_uk  # If you work with Ukrainian text
-
-# Speaker diarization
-pip install diart
-
-# Alternative Whisper backends (default is faster-whisper)
-pip install whisperlivekit[whisper]              # Original Whisper
-pip install whisperlivekit[whisper-timestamped]  # Improved timestamps
-pip install whisperlivekit[mlx-whisper]          # Apple Silicon optimization
-pip install whisperlivekit[openai]               # OpenAI API
-pip install whisperlivekit[simulstreaming]
-```
-
-### 🎹 Pyannote Models Setup
-
-For diarization, you need access to pyannote.audio models:
-
-1. [Accept user conditions](https://huggingface.co/pyannote/segmentation) for the `pyannote/segmentation` model
-2. [Accept user conditions](https://huggingface.co/pyannote/segmentation-3.0) for the `pyannote/segmentation-3.0` model
-3. [Accept user conditions](https://huggingface.co/pyannote/embedding) for the `pyannote/embedding` model
-4. Login with HuggingFace:
-```bash
-pip install huggingface_hub
-huggingface-cli login
-```
+ 
+> **Pyannote Models Setup** For diarization, you need access to pyannote.audio models:
+> 1. [Accept user conditions](https://huggingface.co/pyannote/segmentation) for the `pyannote/segmentation` model
+> 2. [Accept user conditions](https://huggingface.co/pyannote/segmentation-3.0) for the `pyannote/segmentation-3.0` model
+> 3. [Accept user conditions](https://huggingface.co/pyannote/embedding) for the `pyannote/embedding` model
+>4. Login with HuggingFace:
+> ```bash
+> huggingface-cli login
+> ```
 
 ## 💻 Usage Examples
 
-### Command-line Interface
+#### Command-line Interface
 
 Start the transcription server with various options:
 
 ```bash
-# Basic server with English model
-whisperlivekit-server --model tiny.en
+# SimulStreaming backend for ultra-low latency
+whisperlivekit-server --backend simulstreaming --model large-v3
 
 # Advanced configuration with diarization
-whisperlivekit-server --host 0.0.0.0 --port 8000 --model medium --diarization --language auto
-
-# SimulStreaming backend for ultra-low latency
-whisperlivekit-server --backend simulstreaming --model large-v3 --frame-threshold 20
+whisperlivekit-server --host 0.0.0.0 --port 8000 --model medium --diarization --language fr
 ```
 
 
-### Python API Integration (Backend)
-Check [basic_server.py](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/basic_server.py) for a complete example.
+#### Python API Integration (Backend)
+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, get_web_interface_html, parse_args
+from whisperlivekit import TranscriptionEngine, AudioProcessor, parse_args
 from fastapi import FastAPI, WebSocket, WebSocketDisconnect
 from fastapi.responses import HTMLResponse
 from contextlib import asynccontextmanager
 import asyncio
 
-# Global variable for the transcription engine
 transcription_engine = None
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
     global transcription_engine
-    # Example: Initialize with specific parameters directly
-    # You can also load from command-line arguments using parse_args()
-    # args = parse_args()
-    # transcription_engine = TranscriptionEngine(**vars(args))
     transcription_engine = TranscriptionEngine(model="medium", diarization=True, lan="en")
     yield
 
 app = FastAPI(lifespan=lifespan)
 
-# Serve the web interface
-@app.get("/")
-async def get():
-    return HTMLResponse(get_web_interface_html())
-
-# Process WebSocket connections
 async def handle_websocket_results(websocket: WebSocket, results_generator):
-    try:
-        async for response in results_generator:
-            await websocket.send_json(response)
-        await websocket.send_json({"type": "ready_to_stop"})
-    except WebSocketDisconnect:
-        print("WebSocket disconnected during results handling.")
+    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):
@@ -182,64 +131,43 @@ async def websocket_endpoint(websocket: WebSocket):
     # Create a new AudioProcessor for each connection, passing the shared engine
     audio_processor = AudioProcessor(transcription_engine=transcription_engine)    
     results_generator = await audio_processor.create_tasks()
-    send_results_to_client = handle_websocket_results(websocket, results_generator)
-    results_task = asyncio.create_task(send_results_to_client)
+    results_task = asyncio.create_task(handle_websocket_results(websocket, results_generator))
     await websocket.accept()
-    try:
-        while True:
-            message = await websocket.receive_bytes()
-            await audio_processor.process_audio(message)        
-    except WebSocketDisconnect:
-        print(f"Client disconnected: {websocket.client}")
-    except Exception as e:
-        await websocket.close(code=1011, reason=f"Server error: {e}")
-    finally:
-        results_task.cancel()
-        try:
-            await results_task
-        except asyncio.CancelledError:
-            logger.info("Results task successfully cancelled.")
+    while True:
+        message = await websocket.receive_bytes()
+        await audio_processor.process_audio(message)        
 ```
 
-### Frontend Implementation
+#### Frontend Implementation
 
-The package includes a simple HTML/JavaScript implementation that you can adapt for your project. You can find it in `whisperlivekit/web/live_transcription.html`, or load its content using the `get_web_interface_html()` function from `whisperlivekit`:
+The package includes an HTML/JavaScript implementation [here](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/web/live_transcription.html). You can also import it using `from whisperlivekit import get_web_interface_html` & `page = get_web_interface_html()`
 
-```python
-from whisperlivekit import get_web_interface_html
 
-# ... later in your code where you need the HTML string ...
-html_content = get_web_interface_html()
-```
-
-## ⚙️ Configuration Reference
-
-WhisperLiveKit offers extensive configuration options:
+### ⚙️ Parameters & Configuration
 
 | Parameter | Description | Default |
 |-----------|-------------|---------|
-| `--host` | Server host address | `localhost` |
-| `--port` | Server port | `8000` |
-| `--model` | Whisper model size. Caution : '.en' models do not work with Simulstreaming | `tiny` |
+| `--model` | Whisper model size. | `small` |
 | `--language` | Source language code or `auto` | `en` |
 | `--task` | `transcribe` or `translate` | `transcribe` |
-| `--backend` | Processing backend | `faster-whisper` |
-| `--diarization` | Enable speaker identification | `False` |
-| `--punctuation-split` | Use punctuation to improve speaker boundaries | `True` |
-| `--confidence-validation` | Use confidence scores for faster validation | `False` |
+| `--backend` | Processing backend | `simulstreaming` |
 | `--min-chunk-size` | Minimum audio chunk size (seconds) | `1.0` |
-| `--vac` | Use Voice Activity Controller | `False` |
+| `--no-vac` | Disable Voice Activity Controller | `False` |
 | `--no-vad` | Disable Voice Activity Detection | `False` |
-| `--buffer_trimming` | Buffer trimming strategy (`sentence` or `segment`) | `segment` |
 | `--warmup-file` | Audio file path for model warmup | `jfk.wav` |
+| `--host` | Server host address | `localhost` |
+| `--port` | Server port | `8000` |
 | `--ssl-certfile` | Path to the SSL certificate file (for HTTPS support) | `None` |
 | `--ssl-keyfile` | Path to the SSL private key file (for HTTPS support) | `None` |
-| `--segmentation-model` | Hugging Face model ID for pyannote.audio 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 pyannote.audio embedding model. [Available models](https://github.com/juanmc2005/diart/tree/main?tab=readme-ov-file#pre-trained-models) | `speechbrain/spkrec-ecapa-voxceleb` |
 
-**SimulStreaming-specific Options:**
 
-| Parameter | Description | Default |
+| WhisperStreaming backend options | Description | Default |
+|-----------|-------------|---------|
+| `--confidence-validation` | Use confidence scores for faster validation | `False` |
+| `--buffer_trimming` | Buffer trimming strategy (`sentence` or `segment`) | `segment` |
+
+
+| SimulStreaming backend options | Description | Default |
 |-----------|-------------|---------|
 | `--frame-threshold` | AlignAtt frame threshold (lower = faster, higher = more accurate) | `25` |
 | `--beams` | Number of beams for beam search (1 = greedy decoding) | `1` |
@@ -252,115 +180,62 @@ WhisperLiveKit offers extensive configuration options:
 | `--static-init-prompt` | Static prompt that doesn't scroll | `None` |
 | `--max-context-tokens` | Maximum context tokens | `None` |
 | `--model-path` | Direct path to .pt model file. Download it if not found | `./base.pt` |
+| `--preloaded-model-count` | Optional. Number of models to preload in memory to speed up loading (set up to the expected number of concurrent users) | `1` |
 
-## 🔧 How It Works
+| Diarization options | Description | Default |
+|-----------|-------------|---------|
+| `--diarization` | Enable speaker identification | `False` |
+| `--punctuation-split` | Use punctuation to improve speaker boundaries | `True` |
+| `--segmentation-model` | Hugging Face model ID for pyannote.audio 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 pyannote.audio embedding model. [Available models](https://github.com/juanmc2005/diart/tree/main?tab=readme-ov-file#pre-trained-models) | `speechbrain/spkrec-ecapa-voxceleb` |
 
-1. **Audio Capture**: Browser's MediaRecorder API captures audio in webm/opus format
-2. **Streaming**: Audio chunks are sent to the server via WebSocket
-3. **Processing**: Server decodes audio with FFmpeg and streams into Whisper for transcription
-4. **Real-time Output**: 
-   - Partial transcriptions appear immediately in light gray (the 'aperçu')
-   - Finalized text appears in normal color
-   - (When enabled) Different speakers are identified and highlighted
-
-## 🚀 Deployment Guide
+### 🚀 Deployment Guide
 
 To deploy WhisperLiveKit in production:
-
-1. **Server Setup** (Backend):
+ 
+1. **Server Setup**: Install production ASGI server & launch with multiple workers
    ```bash
-   # Install production ASGI server
    pip install uvicorn gunicorn
-
-   # Launch with multiple workers
    gunicorn -k uvicorn.workers.UvicornWorker -w 4 your_app:app
    ```
 
-2. **Frontend Integration**:
-   - Host your customized version of the example HTML/JS in your web application
-   - Ensure WebSocket connection points to your server's address
+2. **Frontend**: Host your customized version of the `html` example & ensure WebSocket connection points correctly
 
 3. **Nginx Configuration** (recommended for production):
-   ```nginx
+    ```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;
-    }
-}
-```
+        location / {
+            proxy_pass http://localhost:8000;
+            proxy_set_header Upgrade $http_upgrade;
+            proxy_set_header Connection "upgrade";
+            proxy_set_header Host $host;
+    }}
+    ```
 
 4. **HTTPS Support**: For secure deployments, use "wss://" instead of "ws://" in WebSocket URL
 
 ### 🐋 Docker
 
-A basic Dockerfile is provided which allows re-use of Python package installation options. See below usage examples:
+A Dockerfile is provided which allows re-use of Python package installation options. Create a reusable image with only the basics and then run as a named container:
 
-**NOTE:** For **larger** models, ensure that your **docker runtime** has enough **memory** available.
-
-#### All defaults
-- Create a reusable image with only the basics and then run as a named container:
 ```bash
 docker build -t whisperlivekit-defaults .
-docker create --gpus all --name whisperlivekit -p 8000:8000 whisperlivekit-defaults
+docker create --gpus all --name whisperlivekit -p 8000:8000 whisperlivekit-defaults --model base
 docker start -i whisperlivekit
 ```
 
+> **Note**: For **large** models, ensure that your **docker runtime** has enough **memory** available
+
 > **Note**: If you're running on a system without NVIDIA GPU support (such as Mac with Apple Silicon or any system without CUDA capabilities), you need to **remove the `--gpus all` flag** from the `docker create` command. Without GPU acceleration, transcription will use CPU only, which may be significantly slower. Consider using small models for better performance on CPU-only systems.
 
 #### Customization
-- Customize the container options:
-```bash
-docker build -t whisperlivekit-defaults .
-docker create --gpus all --name whisperlivekit-base -p 8000:8000 whisperlivekit-defaults --model base
-docker start -i whisperlivekit-base
-```
 
 - `--build-arg` Options:
   - `EXTRAS="whisper-timestamped"` - Add extras to the image's installation (no spaces). Remember to set necessary container options!
   - `HF_PRECACHE_DIR="./.cache/"` - Pre-load a model cache for faster first-time start
-  - `HF_TOKEN="./token"` - Add your Hugging Face Hub access token to download gated models
+  - `HF_TKN_FILE="./token"` - Add your Hugging Face Hub access token to download gated models
 
 ## 🔮 Use Cases
-
-- **Meeting Transcription**: Capture discussions in real-time
-- **Accessibility Tools**: Help hearing-impaired users follow conversations
-- **Content Creation**: Transcribe podcasts or videos automatically
-- **Customer Service**: Transcribe support calls with speaker identification
-
-## 📄 License
-
-This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details.
-
-**⚠️ Important**: When using the SimulStreaming backend, you must also comply with the **PolyForm Noncommercial License 1.0.0** that governs SimulStreaming. For commercial use of the SimulStreaming backend, obtain a commercial license from the [SimulStreaming authors](https://github.com/ufal/SimulStreaming#-licence-and-contributions).
-
-## 🤝 Contributing
-
-Contributions are welcome! Here's how to get started:
-
-1. Fork the repository
-2. Create a feature branch: `git checkout -b feature/amazing-feature`
-3. Commit your changes: `git commit -m 'Add amazing feature'`
-4. Push to your branch: `git push origin feature/amazing-feature`
-5. Open a Pull Request
-
-## 🙏 Acknowledgments
-
-This project builds upon the foundational work of:
-- [Whisper Streaming](https://github.com/ufal/whisper_streaming)
-- [SimulStreaming](https://github.com/ufal/SimulStreaming) (BETA backend)
-- [Diart](https://github.com/juanmc2005/diart)
-- [OpenAI Whisper](https://github.com/openai/whisper)
-
-We extend our gratitude to the original authors for their contributions.
-
-## 🔗 Links
-
-- [GitHub Repository](https://github.com/QuentinFuxa/WhisperLiveKit)
-- [PyPI Package](https://pypi.org/project/whisperlivekit/)
-- [Issue Tracker](https://github.com/QuentinFuxa/WhisperLiveKit/issues)
+Capture discussions in real-time for meeting transcription, help hearing-impaired users follow conversations through accessibility tools, transcribe podcasts or videos automatically for content creation, transcribe support calls with speaker identification for customer service...

pyproject.toml

@@ -0,0 +1,56 @@
+[build-system]
+requires = ["setuptools>=61.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "whisperlivekit"
+version = "0.2.6"
+description = "Real-time, Fully Local Whisper's Speech-to-Text and Speaker Diarization"
+readme = "README.md"
+authors = [
+    { name = "Quentin Fuxa" }
+]
+license = { file = "LICENSE" }
+requires-python = ">=3.9"
+classifiers = [
+    "Development Status :: 4 - Beta",
+    "Intended Audience :: Developers",
+    "License :: OSI Approved :: MIT License",
+    "Programming Language :: Python :: 3.9",
+    "Programming Language :: Python :: 3.10",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+    "Topic :: Multimedia :: Sound/Audio :: Speech"
+]
+dependencies = [
+    "fastapi",
+    "librosa",
+    "soundfile",
+    "faster-whisper",
+    "uvicorn",
+    "websockets",
+    "torch",
+    "tqdm",
+    "tiktoken",
+    'triton>=2.0.0,<3; platform_machine == "x86_64" and (sys_platform == "linux" or sys_platform == "linux2")'
+]
+
+[project.optional-dependencies]
+diarization = ["diart"]
+sentence = ["mosestokenizer", "wtpsplit"]
+whisper = ["whisper"]
+whisper-timestamped = ["whisper-timestamped"]
+mlx-whisper = ["mlx-whisper"]
+openai = ["openai"]
+
+[project.urls]
+Homepage = "https://github.com/QuentinFuxa/WhisperLiveKit"
+
+[project.scripts]
+whisperlivekit-server = "whisperlivekit.basic_server:main"
+
+[tool.setuptools]
+packages = ["whisperlivekit", "whisperlivekit.diarization", "whisperlivekit.simul_whisper", "whisperlivekit.simul_whisper.whisper", "whisperlivekit.simul_whisper.whisper.assets", "whisperlivekit.simul_whisper.whisper.normalizers", "whisperlivekit.web", "whisperlivekit.whisper_streaming_custom"]
+
+[tool.setuptools.package-data]
+whisperlivekit = ["web/*.html", "web/*.css", "web/*.js", "web/src/*.svg"]
+"whisperlivekit.simul_whisper.whisper.assets" = ["*.tiktoken", "*.npz"]

setup.py

@@ -1,54 +0,0 @@
-from setuptools import setup, find_packages
-setup(
-    name="whisperlivekit",
-    version="0.2.1",
-    description="Real-time, Fully Local Whisper's Speech-to-Text and Speaker Diarization",
-    long_description=open("README.md", "r", encoding="utf-8").read(),
-    long_description_content_type="text/markdown",
-    author="Quentin Fuxa",
-    url="https://github.com/QuentinFuxa/WhisperLiveKit",
-    packages=find_packages(),
-    install_requires=[
-        "fastapi",
-        "ffmpeg-python",
-        "librosa",
-        "soundfile",
-        "faster-whisper",
-        "uvicorn",
-        "websockets",
-    ],
-    extras_require={
-        "diarization": ["diart"],
-        "vac": ["torch"],
-        "sentence": ["mosestokenizer", "wtpsplit"],
-        "whisper": ["whisper"],
-        "whisper-timestamped": ["whisper-timestamped"],
-        "mlx-whisper": ["mlx-whisper"],
-        "openai": ["openai"],
-        "simulstreaming": [
-            "torch",
-            "tqdm",
-            "tiktoken",
-            "triton>=2.0.0,<3;platform_machine==\"x86_64\" and sys_platform==\"linux\" or sys_platform==\"linux2\"",
-        ],
-    },
-    package_data={
-        'whisperlivekit': ['web/*.html'],
-        'whisperlivekit.simul_whisper': ['dual_license_simulstreaming.md'],
-    },
-    entry_points={
-        'console_scripts': [
-            'whisperlivekit-server=whisperlivekit.basic_server:main',
-        ],
-    },
-    classifiers=[
-        "Development Status :: 4 - Beta",
-        "Intended Audience :: Developers",
-        "License :: OSI Approved :: MIT License",
-        "Programming Language :: Python :: 3.9",
-        "Programming Language :: Python :: 3.10",
-        "Topic :: Scientific/Engineering :: Artificial Intelligence",
-        "Topic :: Multimedia :: Sound/Audio :: Speech",
-    ],
-    python_requires=">=3.9",
-)

whisperlivekit/__init__.py

@@ -1,5 +1,12 @@
-from .core import TranscriptionEngine
 from .audio_processor import AudioProcessor
-from .web.web_interface import get_web_interface_html
+from .core import TranscriptionEngine
 from .parse_args import parse_args
-__all__ = ['TranscriptionEngine', 'AudioProcessor', 'get_web_interface_html', 'parse_args']
+from .web.web_interface import get_web_interface_html
+
+__all__ = [
+    "TranscriptionEngine",
+    "AudioProcessor",
+    "parse_args",
+    "get_web_interface_html",
+    "download_simulstreaming_backend",
+]

whisperlivekit/audio_processor.py

@@ -1,15 +1,16 @@
 import asyncio
 import numpy as np
-import ffmpeg
 from time import time, sleep
 import math
 import logging
 import traceback
 from datetime import timedelta
-from whisperlivekit.timed_objects import ASRToken
-from whisperlivekit.whisper_streaming_custom.whisper_online import online_factory
-from whisperlivekit.core import TranscriptionEngine
-
+from whisperlivekit.timed_objects import ASRToken, Silence
+from whisperlivekit.core import TranscriptionEngine, online_factory
+from whisperlivekit.ffmpeg_manager import FFmpegManager, FFmpegState
+from whisperlivekit.remove_silences import handle_silences
+from whisperlivekit.trail_repetition import trim_tail_repetition
+from whisperlivekit.silero_vad_iterator import FixedVADIterator
 # Set up logging once
 logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
 logger = logging.getLogger(__name__)
@@ -46,17 +47,19 @@ class AudioProcessor:
         self.last_ffmpeg_activity = time()
         self.ffmpeg_health_check_interval = 5
         self.ffmpeg_max_idle_time = 10
+        self.debug = False
 
         # State management
         self.is_stopping = False
+        self.silence = False
+        self.silence_duration = 0.0
         self.tokens = []
         self.buffer_transcription = ""
         self.buffer_diarization = ""
-        self.full_transcription = ""
         self.end_buffer = 0
         self.end_attributed_speaker = 0
         self.lock = asyncio.Lock()
-        self.beg_loop = time()
+        self.beg_loop = None #to deal with a potential little lag at the websocket initialization, this is now set in process_audio
         self.sep = " "  # Default separator
         self.last_response_content = ""
         
@@ -64,7 +67,24 @@ class AudioProcessor:
         self.asr = models.asr
         self.tokenizer = models.tokenizer
         self.diarization = models.diarization
-        self.ffmpeg_process = self.start_ffmpeg_decoder()
+        self.vac_model = models.vac_model
+        if self.args.vac:
+            self.vac = FixedVADIterator(models.vac_model)
+        else:
+            self.vac = None
+            
+        self.ffmpeg_manager = FFmpegManager(
+            sample_rate=self.sample_rate,
+            channels=self.channels
+        )
+        
+        async def handle_ffmpeg_error(error_type: str):
+            logger.error(f"FFmpeg error: {error_type}")
+            self._ffmpeg_error = error_type
+        
+        self.ffmpeg_manager.on_error_callback = handle_ffmpeg_error
+        self._ffmpeg_error = None
+        
         self.transcription_queue = asyncio.Queue() if self.args.transcription else None
         self.diarization_queue = asyncio.Queue() if self.args.diarization else None
         self.pcm_buffer = bytearray()
@@ -84,90 +104,23 @@ class AudioProcessor:
         """Convert PCM buffer in s16le format to normalized NumPy array."""
         return np.frombuffer(pcm_buffer, dtype=np.int16).astype(np.float32) / 32768.0
 
-    def start_ffmpeg_decoder(self):
-        """Start FFmpeg process for WebM to PCM conversion."""
-        try:
-            return (ffmpeg.input("pipe:0", format="webm")
-                    .output("pipe:1", format="s16le", acodec="pcm_s16le", 
-                            ac=self.channels, ar=str(self.sample_rate))
-                    .run_async(pipe_stdin=True, pipe_stdout=True, pipe_stderr=True))
-        except FileNotFoundError:
-            error = """
-            FFmpeg is not installed or not found in your system's PATH.
-            Please install FFmpeg to enable audio processing.
-
-            Installation instructions:
-
-            # Ubuntu/Debian:
-            sudo apt update && sudo apt install ffmpeg
-
-            # macOS (using Homebrew):
-            brew install ffmpeg
-
-            # Windows:
-            # 1. Download the latest static build from https://ffmpeg.org/download.html
-            # 2. Extract the archive (e.g., to C:\\FFmpeg).
-            # 3. Add the 'bin' directory (e.g., C:\\FFmpeg\\bin) to your system's PATH environment variable.
-
-            After installation, please restart the application.
-            """
-            logger.error(error)
-            raise FileNotFoundError(error)
-
-    async def restart_ffmpeg(self):
-        """Restart the FFmpeg process after failure."""
-        logger.warning("Restarting FFmpeg process...")
-        
-        if self.ffmpeg_process:
-            try:
-                # we check if process is still running
-                if self.ffmpeg_process.poll() is None:
-                    logger.info("Terminating existing FFmpeg process")
-                    self.ffmpeg_process.stdin.close()
-                    self.ffmpeg_process.terminate()
-                    
-                    # wait for termination with timeout
-                    try:
-                        await asyncio.wait_for(
-                            asyncio.get_event_loop().run_in_executor(None, self.ffmpeg_process.wait),
-                            timeout=5.0
-                        )
-                    except asyncio.TimeoutError:
-                        logger.warning("FFmpeg process did not terminate, killing forcefully")
-                        self.ffmpeg_process.kill()
-                        await asyncio.get_event_loop().run_in_executor(None, self.ffmpeg_process.wait)
-            except Exception as e:
-                logger.error(f"Error during FFmpeg process termination: {e}")
-                logger.error(traceback.format_exc())
-        
-        # we start new process
-        try:
-            logger.info("Starting new FFmpeg process")
-            self.ffmpeg_process = self.start_ffmpeg_decoder()
-            self.pcm_buffer = bytearray()
-            self.last_ffmpeg_activity = time()
-            logger.info("FFmpeg process restarted successfully")
-        except Exception as e:
-            logger.error(f"Failed to restart FFmpeg process: {e}")
-            logger.error(traceback.format_exc())
-            # try again after 5s
-            await asyncio.sleep(5)
-            try:
-                self.ffmpeg_process = self.start_ffmpeg_decoder()
-                self.pcm_buffer = bytearray()
-                self.last_ffmpeg_activity = time()
-                logger.info("FFmpeg process restarted successfully on second attempt")
-            except Exception as e2:
-                logger.critical(f"Failed to restart FFmpeg process on second attempt: {e2}")
-                logger.critical(traceback.format_exc())
-
-    async def update_transcription(self, new_tokens, buffer, end_buffer, full_transcription, sep):
+    async def update_transcription(self, new_tokens, buffer, end_buffer, sep):
         """Thread-safe update of transcription with new data."""
         async with self.lock:
             self.tokens.extend(new_tokens)
+            
+            # self.tokens, has_been_trimmed = trim_tail_repetition(
+            #     self.tokens,
+            #     key=lambda t: t.text.strip().lower(),
+            #     min_block=2,        # avoid trimming single '.' loops; set to 1 if you want to remove those too
+            #     max_tail=200,
+            #     prefer="longest",   # prefer removing the longest repeated phrase
+            #     keep=1
+            # )
+            # if has_been_trimmed:
+            #     print('HAS BEEN TRIMMED !')
             self.buffer_transcription = buffer
             self.end_buffer = end_buffer
-            self.full_transcription = full_transcription
             self.sep = sep
             
     async def update_diarization(self, end_attributed_speaker, buffer_diarization=""):
@@ -194,12 +147,12 @@ class AudioProcessor:
             # Calculate remaining times
             remaining_transcription = 0
             if self.end_buffer > 0:
-                remaining_transcription = max(0, round(current_time - self.beg_loop - self.end_buffer, 2))
+                remaining_transcription = max(0, round(current_time - self.beg_loop - self.end_buffer, 1))
                 
             remaining_diarization = 0
             if self.tokens:
                 latest_end = max(self.end_buffer, self.tokens[-1].end if self.tokens else 0)
-                remaining_diarization = max(0, round(latest_end - self.end_attributed_speaker, 2))
+                remaining_diarization = max(0, round(latest_end - self.end_attributed_speaker, 1))
                 
             return {
                 "tokens": self.tokens.copy(),
@@ -218,44 +171,44 @@ class AudioProcessor:
             self.tokens = []
             self.buffer_transcription = self.buffer_diarization = ""
             self.end_buffer = self.end_attributed_speaker = 0
-            self.full_transcription = self.last_response_content = ""
             self.beg_loop = time()
 
     async def ffmpeg_stdout_reader(self):
         """Read audio data from FFmpeg stdout and process it."""
-        loop = asyncio.get_event_loop()
         beg = time()
         
         while True:
             try:
+                # Check if FFmpeg is running
+                state = await self.ffmpeg_manager.get_state()
+                if state == FFmpegState.FAILED:
+                    logger.error("FFmpeg is in FAILED state, cannot read data")
+                    break
+                elif state == FFmpegState.STOPPED:
+                    logger.info("FFmpeg is stopped")
+                    break
+                elif state != FFmpegState.RUNNING:
+                    logger.warning(f"FFmpeg is in {state} state, waiting...")
+                    await asyncio.sleep(0.5)
+                    continue
+                
                 current_time = time()
                 elapsed_time = math.floor((current_time - beg) * 10) / 10
                 buffer_size = max(int(32000 * elapsed_time), 4096)
                 beg = current_time
 
-                # Detect idle state much more quickly
-                if current_time - self.last_ffmpeg_activity > self.ffmpeg_max_idle_time:
-                    logger.warning(f"FFmpeg process idle for {current_time - self.last_ffmpeg_activity:.2f}s. Restarting...")
-                    await self.restart_ffmpeg()
-                    beg = time()
-                    self.last_ffmpeg_activity = time()
-                    continue
-
-                chunk = await loop.run_in_executor(None, self.ffmpeg_process.stdout.read, buffer_size)
-                if chunk:
-                    self.last_ffmpeg_activity = time()
+                chunk = await self.ffmpeg_manager.read_data(buffer_size)
                         
                 if not chunk:
-                    logger.info("FFmpeg stdout closed, no more data to read.")
-                    break
+                    if self.is_stopping:
+                        logger.info("FFmpeg stdout closed, stopping.")
+                        break
+                    else:
+                        # No data available, but not stopping - FFmpeg might be restarting
+                        await asyncio.sleep(0.1)
+                        continue
                     
                 self.pcm_buffer.extend(chunk)
-                        
-                # Send to diarization if enabled
-                if self.args.diarization and self.diarization_queue:
-                    await self.diarization_queue.put(
-                        self.convert_pcm_to_float(self.pcm_buffer).copy()
-                    )
 
                 # Process when enough data
                 if len(self.pcm_buffer) >= self.bytes_per_sec:
@@ -269,18 +222,53 @@ class AudioProcessor:
                     pcm_array = self.convert_pcm_to_float(self.pcm_buffer[:self.max_bytes_per_sec])
                     self.pcm_buffer = self.pcm_buffer[self.max_bytes_per_sec:]
                     
-                    # Send to transcription if enabled
-                    if self.args.transcription and self.transcription_queue:
-                        await self.transcription_queue.put(pcm_array.copy())
+                    res = None
+                    end_of_audio = False
+                    silence_buffer = 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 silence_buffer:
+                        if self.args.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 not self.silence:                            
+                        if self.args.transcription and self.transcription_queue:
+                            await self.transcription_queue.put(pcm_array.copy())
+
+                        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()
+
                     # Sleep if no processing is happening
                     if not self.args.transcription and not self.args.diarization:
                         await asyncio.sleep(0.1)
                     
+                    
+                    
             except Exception as e:
                 logger.warning(f"Exception in ffmpeg_stdout_reader: {e}")
                 logger.warning(f"Traceback: {traceback.format_exc()}")
-                break
+                # Try to recover by waiting a bit
+                await asyncio.sleep(1)
+                
+                # Check if we should exit
+                if self.is_stopping:
+                    break
         
         logger.info("FFmpeg stdout processing finished. Signaling downstream processors.")
         if self.args.transcription and self.transcription_queue:
@@ -293,36 +281,48 @@ class AudioProcessor:
 
     async def transcription_processor(self):
         """Process audio chunks for transcription."""
-        self.full_transcription = ""
         self.sep = self.online.asr.sep
         cumulative_pcm_duration_stream_time = 0.0
         
         while True:
             try:
-                pcm_array = await self.transcription_queue.get()
-                if pcm_array is SENTINEL:
+                item = await self.transcription_queue.get()
+                if item is SENTINEL:
                     logger.debug("Transcription processor received sentinel. Finishing.")
                     self.transcription_queue.task_done()
                     break
                 
-                if not self.online: # Should not happen if queue is used
+                if not self.online:
                     logger.warning("Transcription processor: self.online not initialized.")
                     self.transcription_queue.task_done()
                     continue
 
                 asr_internal_buffer_duration_s = len(getattr(self.online, 'audio_buffer', [])) / self.online.SAMPLING_RATE
                 transcription_lag_s = max(0.0, time() - self.beg_loop - self.end_buffer)
-
-                logger.info(
-                    f"ASR processing: internal_buffer={asr_internal_buffer_duration_s:.2f}s, "
-                    f"lag={transcription_lag_s:.2f}s."
-                )
+                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"
+                    if self.tokens:
+                        asr_processing_logs += " | last_end = {self.tokens[-1].end} |"
+                logger.info(asr_processing_logs)
                 
-                # Process transcription
-                duration_this_chunk = len(pcm_array) / self.sample_rate if isinstance(pcm_array, np.ndarray) else 0
+                if type(item) is Silence:
+                    cumulative_pcm_duration_stream_time += item.duration
+                    self.online.insert_silence(item.duration, self.tokens[-1].end)
+                    continue
+                
+                if isinstance(item, np.ndarray):
+                    pcm_array = item
+                else:
+                    raise Exception('item should be pcm_array')
+                
+                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
 
+                
+                    
+
                 self.online.insert_audio_chunk(pcm_array, stream_time_end_of_current_pcm)
                 new_tokens, current_audio_processed_upto = self.online.process_iter()
                 
@@ -332,8 +332,6 @@ class AudioProcessor:
 
                 if new_tokens:
                     validated_text = self.sep.join([t.text for t in new_tokens])
-                    self.full_transcription += validated_text
-                    
                     if buffer_text.startswith(validated_text):
                         buffer_text = buffer_text[len(validated_text):].lstrip()
 
@@ -350,7 +348,7 @@ class AudioProcessor:
                 new_end_buffer = max(candidate_end_times)
                 
                 await self.update_transcription(
-                    new_tokens, buffer_text, new_end_buffer, self.full_transcription, self.sep
+                    new_tokens, buffer_text, new_end_buffer, self.sep
                 )
                 self.transcription_queue.task_done()
                 
@@ -365,25 +363,35 @@ class AudioProcessor:
     async def diarization_processor(self, diarization_obj):
         """Process audio chunks for speaker diarization."""
         buffer_diarization = ""
-        
+        cumulative_pcm_duration_stream_time = 0.0
         while True:
             try:
-                pcm_array = await self.diarization_queue.get()
-                if pcm_array is SENTINEL:
+                item = await self.diarization_queue.get()
+                if item is SENTINEL:
                     logger.debug("Diarization processor received sentinel. Finishing.")
                     self.diarization_queue.task_done()
                     break
                 
+                if type(item) is Silence:
+                    cumulative_pcm_duration_stream_time += item.duration
+                    diarization_obj.insert_silence(item.duration)
+                    continue
+    
+                if isinstance(item, np.ndarray):
+                    pcm_array = item
+                else:
+                    raise Exception('item should be pcm_array') 
+                
                 # Process diarization
                 await diarization_obj.diarize(pcm_array)
                 
                 async with self.lock:
-                    new_end = diarization_obj.assign_speakers_to_tokens(
-                        self.end_attributed_speaker, 
+                    self.tokens = diarization_obj.assign_speakers_to_tokens(
                         self.tokens,
                         use_punctuation_split=self.args.punctuation_split
                     )
-                    self.end_attributed_speaker = new_end
+                    if len(self.tokens) > 0:
+                        self.end_attributed_speaker = max(self.tokens[-1].end, self.end_attributed_speaker)
                     if buffer_diarization:
                         self.buffer_diarization = buffer_diarization
                 
@@ -399,8 +407,25 @@ class AudioProcessor:
 
     async def results_formatter(self):
         """Format processing results for output."""
+        last_sent_trans = None
+        last_sent_diar = None
         while True:
             try:
+                ffmpeg_state = await self.ffmpeg_manager.get_state()
+                if ffmpeg_state == FFmpegState.FAILED and self._ffmpeg_error:
+                    yield {
+                        "status": "error",
+                        "error": f"FFmpeg error: {self._ffmpeg_error}",
+                        "lines": [],
+                        "buffer_transcription": "",
+                        "buffer_diarization": "",
+                        "remaining_time_transcription": 0,
+                        "remaining_time_diarization": 0
+                    }
+                    self._ffmpeg_error = None
+                    await asyncio.sleep(1)
+                    continue
+                
                 # Get current state
                 state = await self.get_current_state()
                 tokens = state["tokens"]
@@ -421,13 +446,16 @@ class AudioProcessor:
                 lines = []
                 last_end_diarized = 0
                 undiarized_text = []
-                
-                # Process each token
+                current_time = time() - self.beg_loop if self.beg_loop else None
+                tokens, buffer_transcription, buffer_diarization = handle_silences(tokens, buffer_transcription, buffer_diarization, current_time, self.silence)
                 for token in tokens:
                     speaker = token.speaker
                     
+                    if speaker == -1: #Speaker -1 means no attributed by diarization. In the frontend, it should appear under 'Speaker 1'
+                        speaker = 1
+                    
                     # Handle diarization
-                    if self.args.diarization:
+                    if self.args.diarization and not tokens[-1].speaker == -2:
                         if (speaker in [-1, 0]) and token.end >= end_attributed_speaker:
                             undiarized_text.append(token.text)
                             continue
@@ -436,21 +464,23 @@ class AudioProcessor:
                         if speaker not in [-1, 0]:
                             last_end_diarized = max(token.end, last_end_diarized)
 
-                    # Group by speaker
+                    debug_info = ""
+                    if self.debug:
+                        debug_info = f"[{format_time(token.start)} : {format_time(token.end)}]"
                     if speaker != previous_speaker or not lines:
                         lines.append({
                             "speaker": speaker,
-                            "text": token.text,
+                            "text": token.text + debug_info,
                             "beg": format_time(token.start),
                             "end": format_time(token.end),
                             "diff": round(token.end - last_end_diarized, 2)
                         })
                         previous_speaker = speaker
                     elif token.text:  # Only append if text isn't empty
-                        lines[-1]["text"] += sep + token.text
+                        lines[-1]["text"] += sep + token.text + debug_info
                         lines[-1]["end"] = format_time(token.end)
                         lines[-1]["diff"] = round(token.end - last_end_diarized, 2)
-                
+
                 # Handle undiarized text
                 if undiarized_text:
                     combined = sep.join(undiarized_text)
@@ -487,10 +517,19 @@ class AudioProcessor:
                                            ' '.join([f"{line['speaker']} {line['text']}" for line in final_lines_for_response]) + \
                                            f" | {buffer_transcription} | {buffer_diarization}"
                 
-                if current_response_signature != self.last_response_content and \
-                   (final_lines_for_response or buffer_transcription or buffer_diarization or response_status == "no_audio_detected"):
+                trans = state["remaining_time_transcription"]
+                diar = state["remaining_time_diarization"]
+                should_push = (
+                    current_response_signature != self.last_response_content
+                    or last_sent_trans is None
+                    or round(trans, 1) != round(last_sent_trans, 1)
+                    or round(diar, 1) != round(last_sent_diar, 1)
+                )
+                if should_push and (final_lines_for_response or buffer_transcription or buffer_diarization or response_status == "no_audio_detected" or trans > 0 or diar > 0):
                     yield response
                     self.last_response_content = current_response_signature
+                    last_sent_trans = trans
+                    last_sent_diar = diar
                 
                 # Check for termination condition
                 if self.is_stopping:
@@ -517,6 +556,21 @@ class AudioProcessor:
         self.all_tasks_for_cleanup = []
         processing_tasks_for_watchdog = []
 
+        success = await self.ffmpeg_manager.start()
+        if not success:
+            logger.error("Failed to start FFmpeg manager")
+            async def error_generator():
+                yield {
+                    "status": "error", 
+                    "error": "FFmpeg failed to start. Please check that FFmpeg is installed.",
+                    "lines": [],
+                    "buffer_transcription": "",
+                    "buffer_diarization": "",
+                    "remaining_time_transcription": 0,
+                    "remaining_time_diarization": 0
+                }
+            return error_generator()
+
         if self.args.transcription and self.online:
             self.transcription_task = asyncio.create_task(self.transcription_processor())
             self.all_tasks_for_cleanup.append(self.transcription_task)
@@ -542,8 +596,7 @@ class AudioProcessor:
         while True:
             try:
                 await asyncio.sleep(10)
-                current_time = time()
-
+                
                 for i, task in enumerate(tasks_to_monitor):
                     if task.done():
                         exc = task.exception()
@@ -553,12 +606,15 @@ class AudioProcessor:
                         else:
                             logger.info(f"{task_name} completed normally.")
                 
-                ffmpeg_idle_time = current_time - self.last_ffmpeg_activity
-                if ffmpeg_idle_time > 15:
-                    logger.warning(f"FFmpeg idle for {ffmpeg_idle_time:.2f}s - may need attention.")
-                    if ffmpeg_idle_time > 30 and not self.is_stopping:
-                        logger.error("FFmpeg idle for too long and not in stopping phase, forcing restart.")
-                        await self.restart_ffmpeg()
+                # Check FFmpeg status through the manager
+                ffmpeg_state = await self.ffmpeg_manager.get_state()
+                if ffmpeg_state == FFmpegState.FAILED:
+                    logger.error("FFmpeg is in FAILED state, notifying results formatter")
+                    # FFmpeg manager will handle its own recovery
+                elif ffmpeg_state == FFmpegState.STOPPED and not self.is_stopping:
+                    logger.warning("FFmpeg unexpectedly stopped, attempting restart")
+                    await self.ffmpeg_manager.restart()
+                    
             except asyncio.CancelledError:
                 logger.info("Watchdog task cancelled.")
                 break
@@ -567,7 +623,7 @@ class AudioProcessor:
         
     async def cleanup(self):
         """Clean up resources when processing is complete."""
-        logger.info("Starting cleanup of AudioProcessor resources.")
+        logger.info("Starting cleanup of AudioProcessor resources.")        
         for task in self.all_tasks_for_cleanup:
             if task and not task.done():
                 task.cancel()
@@ -576,26 +632,8 @@ class AudioProcessor:
         if created_tasks:
             await asyncio.gather(*created_tasks, return_exceptions=True)
         logger.info("All processing tasks cancelled or finished.")
-
-        if self.ffmpeg_process:
-            if self.ffmpeg_process.stdin and not self.ffmpeg_process.stdin.closed:
-                try:
-                    self.ffmpeg_process.stdin.close()
-                except Exception as e:
-                    logger.warning(f"Error closing ffmpeg stdin during cleanup: {e}")
-            
-            # Wait for ffmpeg process to terminate
-            if self.ffmpeg_process.poll() is None: # Check if process is still running
-                logger.info("Waiting for FFmpeg process to terminate...")
-                try:
-                    # Run wait in executor to avoid blocking async loop
-                    await asyncio.get_event_loop().run_in_executor(None, self.ffmpeg_process.wait, 5.0) # 5s timeout
-                except Exception as e: # subprocess.TimeoutExpired is not directly caught by asyncio.wait_for with run_in_executor
-                    logger.warning(f"FFmpeg did not terminate gracefully, killing. Error: {e}")
-                    self.ffmpeg_process.kill()
-                    await asyncio.get_event_loop().run_in_executor(None, self.ffmpeg_process.wait) # Wait for kill
-            logger.info("FFmpeg process terminated.")
-
+        await self.ffmpeg_manager.stop()
+        logger.info("FFmpeg manager stopped.")
         if self.args.diarization and hasattr(self, 'diarization') and hasattr(self.diarization, 'close'):
             self.diarization.close()
         logger.info("AudioProcessor cleanup complete.")
@@ -603,77 +641,25 @@ class AudioProcessor:
 
     async def process_audio(self, message):
         """Process incoming audio data."""
-        # If already stopping or stdin is closed, ignore further audio, especially residual chunks.
-        if self.is_stopping or (self.ffmpeg_process and self.ffmpeg_process.stdin and self.ffmpeg_process.stdin.closed):
-            logger.warning(f"AudioProcessor is stopping or stdin is closed. Ignoring incoming audio message (length: {len(message)}).")
-            if not message and self.ffmpeg_process and self.ffmpeg_process.stdin and not self.ffmpeg_process.stdin.closed:
-                 logger.info("Received empty message while already in stopping state; ensuring stdin is closed.")
-                 try:
-                    self.ffmpeg_process.stdin.close()
-                 except Exception as e:
-                    logger.warning(f"Error closing ffmpeg stdin on redundant stop signal during stopping state: {e}")
-            return
 
-        if not message: # primary signal to start stopping
+        if not self.beg_loop:
+            self.beg_loop = time()
+
+        if not message:
             logger.info("Empty audio message received, initiating stop sequence.")
             self.is_stopping = True
-            if self.ffmpeg_process and self.ffmpeg_process.stdin and not self.ffmpeg_process.stdin.closed:
-                try:
-                    self.ffmpeg_process.stdin.close()
-                    logger.info("FFmpeg stdin closed due to primary stop signal.")
-                except Exception as e:
-                    logger.warning(f"Error closing ffmpeg stdin on stop: {e}")
+            # Signal FFmpeg manager to stop accepting data
+            await self.ffmpeg_manager.stop()
             return
 
-        retry_count = 0
-        max_retries = 3
-        
-        # Log periodic heartbeats showing ongoing audio proc
-        current_time = time()
-        if not hasattr(self, '_last_heartbeat') or current_time - self._last_heartbeat >= 10:
-            logger.debug(f"Processing audio chunk, last FFmpeg activity: {current_time - self.last_ffmpeg_activity:.2f}s ago")
-            self._last_heartbeat = current_time
-        
-        while retry_count < max_retries:
-            try:
-                if not self.ffmpeg_process or not hasattr(self.ffmpeg_process, 'stdin') or self.ffmpeg_process.poll() is not None:
-                    logger.warning("FFmpeg process not available, restarting...")
-                    await self.restart_ffmpeg()
-                
-                loop = asyncio.get_running_loop()                
-                try:
-                    await asyncio.wait_for(
-                        loop.run_in_executor(None, lambda: self.ffmpeg_process.stdin.write(message)),
-                        timeout=2.0
-                    )
-                except asyncio.TimeoutError:
-                    logger.warning("FFmpeg write operation timed out, restarting...")
-                    await self.restart_ffmpeg()
-                    retry_count += 1
-                    continue
-                    
-                try:
-                    await asyncio.wait_for(
-                        loop.run_in_executor(None, self.ffmpeg_process.stdin.flush),
-                        timeout=2.0
-                    )
-                except asyncio.TimeoutError:
-                    logger.warning("FFmpeg flush operation timed out, restarting...")
-                    await self.restart_ffmpeg()
-                    retry_count += 1
-                    continue
-                    
-                self.last_ffmpeg_activity = time()
-                return
-                    
-            except (BrokenPipeError, AttributeError, OSError) as e:
-                retry_count += 1
-                logger.warning(f"Error writing to FFmpeg: {e}. Retry {retry_count}/{max_retries}...")
-                
-                if retry_count < max_retries:
-                    await self.restart_ffmpeg()
-                    await asyncio.sleep(0.5)
-                else:
-                    logger.error("Maximum retries reached for FFmpeg process")
-                    await self.restart_ffmpeg()
-                    return
+        if self.is_stopping:
+            logger.warning("AudioProcessor is stopping. Ignoring incoming audio.")
+            return
+
+        success = await self.ffmpeg_manager.write_data(message)
+        if not success:
+            ffmpeg_state = await self.ffmpeg_manager.get_state()
+            if ffmpeg_state == FFmpegState.FAILED:
+                logger.error("FFmpeg is in FAILED state, cannot process audio")
+            else:
+                logger.warning("Failed to write audio data to FFmpeg")

whisperlivekit/basic_server.py

@@ -5,6 +5,9 @@ from fastapi.middleware.cors import CORSMiddleware
 from whisperlivekit import TranscriptionEngine, AudioProcessor, get_web_interface_html, parse_args
 import asyncio
 import logging
+from starlette.staticfiles import StaticFiles
+import pathlib
+import whisperlivekit.web as webpkg
 
 logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
 logging.getLogger().setLevel(logging.WARNING)
@@ -30,6 +33,8 @@ app.add_middleware(
     allow_methods=["*"],
     allow_headers=["*"],
 )
+web_dir = pathlib.Path(webpkg.__file__).parent
+app.mount("/web", StaticFiles(directory=str(web_dir)), name="web")
 
 @app.get("/")
 async def get():
@@ -47,7 +52,7 @@ async def handle_websocket_results(websocket, results_generator):
     except WebSocketDisconnect:
         logger.info("WebSocket disconnected while handling results (client likely closed connection).")
     except Exception as e:
-        logger.warning(f"Error in WebSocket results handler: {e}")
+        logger.error(f"Error in WebSocket results handler: {e}")
 
 
 @app.websocket("/asr")

whisperlivekit/core.py

@@ -1,9 +1,12 @@
 try:
-    from whisperlivekit.whisper_streaming_custom.whisper_online import backend_factory, warmup_asr
+    from whisperlivekit.whisper_streaming_custom.whisper_online import backend_factory
+    from whisperlivekit.whisper_streaming_custom.online_asr import OnlineASRProcessor
 except ImportError:
-    from .whisper_streaming_custom.whisper_online import backend_factory, warmup_asr
+    from .whisper_streaming_custom.whisper_online import backend_factory
+    from .whisper_streaming_custom.online_asr import OnlineASRProcessor
+from whisperlivekit.warmup import warmup_asr, warmup_online
 from argparse import Namespace
-
+import sys
 
 class TranscriptionEngine:
     _instance = None
@@ -22,7 +25,6 @@ class TranscriptionEngine:
             "host": "localhost",
             "port": 8000,
             "warmup_file": None,
-            "confidence_validation": False,
             "diarization": False,
             "punctuation_split": False,
             "min_chunk_size": 0.5,
@@ -32,22 +34,22 @@ class TranscriptionEngine:
             "lan": "auto",
             "task": "transcribe",
             "backend": "faster-whisper",
-            "vac": False,
+            "vac": True,
             "vac_chunk_size": 0.04,
-            "buffer_trimming": "segment",
-            "buffer_trimming_sec": 15,
             "log_level": "DEBUG",
             "ssl_certfile": None,
             "ssl_keyfile": None,
             "transcription": True,
             "vad": True,
-            "segmentation_model": "pyannote/segmentation-3.0",
-            "embedding_model": "pyannote/embedding",
+            # whisperstreaming params:
+            "buffer_trimming": "segment",
+            "confidence_validation": False,
+            "buffer_trimming_sec": 15,
             # simulstreaming params:
             "frame_threshold": 25,
             "beams": 1,
             "decoder_type": None,
-            "audio_max_len": 30.0,
+            "audio_max_len": 20.0,
             "audio_min_len": 0.0,
             "cif_ckpt_path": None,
             "never_fire": False,
@@ -55,6 +57,10 @@ class TranscriptionEngine:
             "static_init_prompt": None,
             "max_context_tokens": None,
             "model_path": './base.pt',
+            "diarization_backend": "diart",
+            # diart params:
+            "segmentation_model": "pyannote/segmentation-3.0",
+            "embedding_model": "pyannote/embedding",
         }
 
         config_dict = {**defaults, **kwargs}
@@ -63,6 +69,8 @@ class TranscriptionEngine:
             config_dict['transcription'] = not kwargs['no_transcription']
         if 'no_vad' in kwargs:
             config_dict['vad'] = not kwargs['no_vad']
+        if 'no_vac' in kwargs:
+            config_dict['vac'] = not kwargs['no_vac']
         
         config_dict.pop('no_transcription', None)
         config_dict.pop('no_vad', None)
@@ -76,17 +84,72 @@ class TranscriptionEngine:
         self.asr = None
         self.tokenizer = None
         self.diarization = None
+        self.vac_model = None
+        
+        if self.args.vac:
+            import torch
+            self.vac_model, _ = torch.hub.load(repo_or_dir="snakers4/silero-vad", model="silero_vad")            
         
         if self.args.transcription:
-            self.asr, self.tokenizer = backend_factory(self.args)
-            warmup_asr(self.asr, self.args.warmup_file)
+            if self.args.backend == "simulstreaming": 
+                from whisperlivekit.simul_whisper import SimulStreamingASR
+                self.tokenizer = None
+                simulstreaming_kwargs = {}
+                for attr in ['frame_threshold', 'beams', 'decoder_type', 'audio_max_len', 'audio_min_len', 
+                            'cif_ckpt_path', 'never_fire', 'init_prompt', 'static_init_prompt', 
+                            'max_context_tokens', 'model_path', 'warmup_file', 'preload_model_count']:
+                    if hasattr(self.args, attr):
+                        simulstreaming_kwargs[attr] = getattr(self.args, attr)
+        
+                # Add segment_length from min_chunk_size
+                simulstreaming_kwargs['segment_length'] = getattr(self.args, 'min_chunk_size', 0.5)
+                simulstreaming_kwargs['task'] = self.args.task
+                
+                size = self.args.model
+                self.asr = SimulStreamingASR(
+                    modelsize=size,
+                    lan=self.args.lan,
+                    cache_dir=getattr(self.args, 'model_cache_dir', None),
+                    model_dir=getattr(self.args, 'model_dir', None),
+                    **simulstreaming_kwargs
+                )
+
+            else:
+                self.asr, self.tokenizer = backend_factory(self.args)
+            warmup_asr(self.asr, self.args.warmup_file) #for simulstreaming, warmup should be done in the online class not here
 
         if self.args.diarization:
-            from whisperlivekit.diarization.diarization_online import DiartDiarization
-            self.diarization = DiartDiarization(
-                block_duration=self.args.min_chunk_size,
-                segmentation_model_name=self.args.segmentation_model,
-                embedding_model_name=self.args.embedding_model
-            )
+            if self.args.diarization_backend == "diart":
+                from whisperlivekit.diarization.diart_backend import DiartDiarization
+                self.diarization = DiartDiarization(
+                    block_duration=self.args.min_chunk_size,
+                    segmentation_model_name=self.args.segmentation_model,
+                    embedding_model_name=self.args.embedding_model
+                )
+            elif self.args.diarization_backend == "sortformer":
+                raise ValueError('Sortformer backend in developement')
+            else:
+                raise ValueError(f"Unknown diarization backend: {self.args.diarization_backend}")
             
         TranscriptionEngine._initialized = True
+
+
+
+def online_factory(args, asr, tokenizer, logfile=sys.stderr):
+    if args.backend == "simulstreaming":    
+        from whisperlivekit.simul_whisper import SimulStreamingOnlineProcessor
+        online = SimulStreamingOnlineProcessor(
+            asr,
+            logfile=logfile,
+        )
+        # warmup_online(online, args.warmup_file)
+    else:
+        online = OnlineASRProcessor(
+            asr,
+            tokenizer,
+            logfile=logfile,
+            buffer_trimming=(args.buffer_trimming, args.buffer_trimming_sec),
+            confidence_validation = args.confidence_validation
+        )
+    return online
+  

whisperlivekit/diarization/diart_backend.py

@@ -29,6 +29,7 @@ class DiarizationObserver(Observer):
         self.speaker_segments = []
         self.processed_time = 0
         self.segment_lock = threading.Lock()
+        self.global_time_offset = 0.0
     
     def on_next(self, value: Tuple[Annotation, Any]):
         annotation, audio = value
@@ -49,8 +50,8 @@ class DiarizationObserver(Observer):
                         print(f"  {speaker}: {start:.2f}s-{end:.2f}s")
                         self.speaker_segments.append(SpeakerSegment(
                             speaker=speaker,
-                            start=start,
-                            end=end
+                            start=start + self.global_time_offset,
+                            end=end + self.global_time_offset
                         ))
             else:
                 logger.debug("\nNo speakers detected in this segment")
@@ -165,7 +166,7 @@ class WebSocketAudioSource(AudioSource):
 
 
 class DiartDiarization:
-    def __init__(self, sample_rate: int = 16000, config : SpeakerDiarizationConfig = None, use_microphone: bool = False, block_duration: float = 0.5, segmentation_model_name: str = "pyannote/segmentation-3.0", embedding_model_name: str = "speechbrain/spkrec-ecapa-voxceleb"):
+    def __init__(self, sample_rate: int = 16000, config : SpeakerDiarizationConfig = None, use_microphone: bool = False, block_duration: float = 1.5, segmentation_model_name: str = "pyannote/segmentation-3.0", embedding_model_name: str = "pyannote/embedding"):
         segmentation_model = m.SegmentationModel.from_pretrained(segmentation_model_name)
         embedding_model = m.EmbeddingModel.from_pretrained(embedding_model_name)
         
@@ -199,6 +200,9 @@ class DiartDiarization:
         self.inference.attach_observers(self.observer)
         asyncio.get_event_loop().run_in_executor(None, self.inference)
 
+    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.
@@ -206,15 +210,14 @@ class DiartDiarization:
         """
         if self.custom_source:
             self.custom_source.push_audio(pcm_array)            
-        self.observer.clear_old_segments()        
-        return self.observer.get_segments()
+        # self.observer.clear_old_segments()        
 
     def close(self):
         """Close the audio source."""
         if self.custom_source:
             self.custom_source.close()
 
-    def assign_speakers_to_tokens(self, end_attributed_speaker, tokens: list, use_punctuation_split: bool = False) -> float:
+    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.
@@ -231,85 +234,82 @@ class DiartDiarization:
         
         if not self.lag_diart and segments and tokens:
             self.lag_diart = segments[0].start - tokens[0].start
-        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
-                    end_attributed_speaker = max(token.end, end_attributed_speaker)
         
-        if use_punctuation_split and len(tokens) > 1:
-            punctuation_marks = {'.', '!', '?'}
-            
-            print("Here are the tokens:", 
-                  [(t.text, t.start, t.end, t.speaker) for t in tokens[:10]])
-            
-            segment_map = []
-            for segment in segments:
-                speaker_num = extract_number(segment.speaker) + 1
-                segment_map.append((segment.start, segment.end, speaker_num))
-            segment_map.sort(key=lambda x: x[0])
-            
-            i = 0
-            while i < len(tokens):
-                current_token = tokens[i]
-                
-                is_sentence_end = False
-                if current_token.text and current_token.text.strip():
-                    text = current_token.text.strip()
-                    if text[-1] in punctuation_marks:
-                        is_sentence_end = True
-                        logger.debug(f"Token {i} ends sentence: '{current_token.text}' at {current_token.end:.2f}s")
-                
-                if is_sentence_end and current_token.speaker != -1:
-                    punctuation_time = current_token.end
-                    current_speaker = current_token.speaker
-                    
-                    j = i + 1
-                    next_sentence_tokens = []                    
-                    while j < len(tokens):
-                        next_token = tokens[j]
-                        next_sentence_tokens.append(j)
-                        
-                        # Check if this token ends the next sentence
-                        if next_token.text and next_token.text.strip():
-                            if next_token.text.strip()[-1] in punctuation_marks:
-                                break
-                        j += 1
-                    
-                    if next_sentence_tokens:
-                        speaker_times = {}
-                        
-                        for idx in next_sentence_tokens:
-                            token = tokens[idx]
-                            # Find which segments overlap with this token
-                            for seg_start, seg_end, seg_speaker in segment_map:
-                                if not (seg_end <= token.start or seg_start >= token.end):
-                                    # Calculate overlap duration
-                                    overlap_start = max(seg_start, token.start)
-                                    overlap_end = min(seg_end, token.end)
-                                    overlap_duration = overlap_end - overlap_start
-                                    
-                                    if seg_speaker not in speaker_times:
-                                        speaker_times[seg_speaker] = 0
-                                    speaker_times[seg_speaker] += overlap_duration
-                        
-                        if speaker_times:
-                            dominant_speaker = max(speaker_times.items(), key=lambda x: x[1])[0]
-                            
-                            if dominant_speaker != current_speaker:
-                                logger.debug(f"  Speaker change after punctuation: {current_speaker} → {dominant_speaker}")
-                                
-                                for idx in next_sentence_tokens:
-                                    if tokens[idx].speaker != dominant_speaker:
-                                        logger.debug(f"    Reassigning token {idx} ('{tokens[idx].text}') to Speaker {dominant_speaker}")
-                                        tokens[idx].speaker = dominant_speaker
-                                        end_attributed_speaker = max(tokens[idx].end, end_attributed_speaker)
-                            else:
-                                for idx in next_sentence_tokens:
-                                    if tokens[idx].speaker == -1:
-                                        tokens[idx].speaker = current_speaker
-                                        end_attributed_speaker = max(tokens[idx].end, end_attributed_speaker)
-                
-                i += 1         
+        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
         
-        return end_attributed_speaker
+def concatenate_speakers(segments):
+    segments_concatenated = [{"speaker": 1, "begin": 0.0, "end": 0.0}]
+    for segment in segments:
+        speaker = extract_number(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
+    # print("Segments concatenated:")
+    # for entry in segments_concatenated:
+    #     print(f"Speaker {entry['speaker']}: {entry['begin']:.2f}s - {entry['end']:.2f}s")   
+    return segments_concatenated
+
+
+def add_speaker_to_tokens(segments, tokens):
+    """
+    Assign speakers to tokens based on diarization segments, with punctuation-aware boundary adjustment.
+    """
+    punctuation_marks = {'.', '!', '?'}
+    punctuation_tokens = [token for token in tokens if token.text.strip() in punctuation_marks]
+    segments_concatenated = concatenate_speakers(segments)
+    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 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 < len(punctuation_tokens) - 1 and ind - 1 >= 0:
+                            segments_concatenated[ind - 1]['begin'] = punctuation_tokens[i - 1].end
+                    break
+
+    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
+
+    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
+                # print(
+                #     f"Token '{token.text}' ('begin': {token.start:.2f}, 'end': {token.end:.2f}) "
+                #     f"assigned to Speaker {segment['speaker']} ('segment': {segment['begin']:.2f}-{segment['end']:.2f})"
+                # )
+            elif token.start > segment['end']:
+                break
+    return tokens
+
+
+def visualize_tokens(tokens):
+    conversation = [{"speaker": -1, "text": ""}]
+    for token in tokens:
+        speaker = conversation[-1]['speaker']
+        if token.speaker != speaker:
+            conversation.append({"speaker": token.speaker, "text": token.text})
+        else:
+            conversation[-1]['text'] += token.text
+    print("Conversation:")
+    for entry in conversation:
+        print(f"Speaker {entry['speaker']}: {entry['text']}")

whisperlivekit/diarization/sortformer_backend.py

@@ -0,0 +1,145 @@
+import numpy as np
+import torch
+import logging
+from whisperlivekit.timed_objects import SpeakerSegment
+
+logger = logging.getLogger(__name__)
+
+try:
+    from nemo.collections.asr.models import SortformerEncLabelModel
+except ImportError:
+    raise SystemExit("""Please use `pip install "git+https://github.com/NVIDIA/NeMo.git@main#egg=nemo_toolkit[asr]"` to use the Sortformer diarization""")
+
+class SortformerDiarization:
+    def __init__(self, model_name="nvidia/diar_streaming_sortformer_4spk-v2"):
+        self.diar_model = SortformerEncLabelModel.from_pretrained(model_name)
+        self.diar_model.eval()
+
+        if torch.cuda.is_available():
+            self.diar_model.to(torch.device("cuda"))
+
+        # Streaming parameters for speed
+        self.diar_model.sortformer_modules.chunk_len = 12
+        self.diar_model.sortformer_modules.chunk_right_context = 1
+        self.diar_model.sortformer_modules.spkcache_len = 188
+        self.diar_model.sortformer_modules.fifo_len = 188
+        self.diar_model.sortformer_modules.spkcache_update_period = 144
+        self.diar_model.sortformer_modules.log = False
+        self.diar_model.sortformer_modules._check_streaming_parameters()
+
+        self.batch_size = 1
+        self.processed_signal_offset = torch.zeros((self.batch_size,), dtype=torch.long, device=self.diar_model.device)
+        
+        self.audio_buffer = np.array([], dtype=np.float32)
+        self.sample_rate = 16000
+        self.speaker_segments = []
+
+        self.streaming_state = self.diar_model.sortformer_modules.init_streaming_state(
+            batch_size=self.batch_size,
+            async_streaming=True,
+            device=self.diar_model.device
+        )
+        self.total_preds = torch.zeros((self.batch_size, 0, self.diar_model.sortformer_modules.n_spk), device=self.diar_model.device)
+
+
+    def _prepare_audio_signal(self, signal):
+        audio_signal = torch.tensor(signal).unsqueeze(0).to(self.diar_model.device)
+        audio_signal_length = torch.tensor([audio_signal.shape[1]]).to(self.diar_model.device)
+        processed_signal, processed_signal_length = self.diar_model.preprocessor(input_signal=audio_signal, length=audio_signal_length)
+        return processed_signal, processed_signal_length
+
+    def _create_streaming_loader(self, processed_signal, processed_signal_length):
+        streaming_loader = self.diar_model.sortformer_modules.streaming_feat_loader(
+            feat_seq=processed_signal,
+            feat_seq_length=processed_signal_length,
+            feat_seq_offset=self.processed_signal_offset,
+        )
+        return streaming_loader
+
+    async def diarize(self, pcm_array: np.ndarray):
+        """
+        Process an incoming audio chunk for diarization.
+        """
+        self.audio_buffer = np.concatenate([self.audio_buffer, pcm_array])
+        
+        # Process in fixed-size chunks (e.g., 1 second)
+        chunk_size = self.sample_rate # 1 second of audio
+        
+        while len(self.audio_buffer) >= chunk_size:
+            chunk_to_process = self.audio_buffer[:chunk_size]
+            self.audio_buffer = self.audio_buffer[chunk_size:]
+
+            processed_signal, processed_signal_length = self._prepare_audio_signal(chunk_to_process)
+            
+            current_offset_seconds = self.processed_signal_offset.item() * self.diar_model.preprocessor._cfg.window_stride
+
+            streaming_loader = self._create_streaming_loader(processed_signal, processed_signal_length)
+            
+            frame_duration_s = self.diar_model.sortformer_modules.subsampling_factor * self.diar_model.preprocessor._cfg.window_stride
+            chunk_duration_seconds = self.diar_model.sortformer_modules.chunk_len * frame_duration_s
+
+            for i, chunk_feat_seq_t, feat_lengths, left_offset, right_offset in streaming_loader:
+                with torch.inference_mode():
+                    self.streaming_state, self.total_preds = self.diar_model.forward_streaming_step(
+                        processed_signal=chunk_feat_seq_t,
+                        processed_signal_length=feat_lengths,
+                        streaming_state=self.streaming_state,
+                        total_preds=self.total_preds,
+                        left_offset=left_offset,
+                        right_offset=right_offset,
+                    )
+                    
+                    num_new_frames = feat_lengths[0].item()
+                    
+                    # Get predictions for the current chunk from the end of total_preds
+                    preds_np = self.total_preds[0, -num_new_frames:].cpu().numpy()
+                    active_speakers = np.argmax(preds_np, axis=1)
+
+                    for idx, spk in enumerate(active_speakers):
+                        start_time = current_offset_seconds + (i * chunk_duration_seconds) + (idx * frame_duration_s)
+                        end_time = start_time + frame_duration_s
+                        
+                        if self.speaker_segments and self.speaker_segments[-1].speaker == spk + 1:
+                            self.speaker_segments[-1].end = end_time
+                        else:
+                            self.speaker_segments.append(SpeakerSegment(
+                                speaker=int(spk + 1),
+                                start=start_time,
+                                end=end_time
+                            ))
+            
+            self.processed_signal_offset += processed_signal_length
+
+
+    def assign_speakers_to_tokens(self, tokens: list, **kwargs) -> list:
+        """
+        Assign speakers to tokens based on timing overlap with speaker segments.
+        """
+        for token in tokens:
+            for segment in self.speaker_segments:
+                if not (segment.end <= token.start or segment.start >= token.end):
+                    token.speaker = segment.speaker
+        return tokens
+
+    def close(self):
+        """
+        Cleanup resources.
+        """
+        logger.info("Closing SortformerDiarization.")
+
+if __name__ == '__main__':
+    import librosa
+    an4_audio = 'new_audio_test.mp3'
+    signal, sr = librosa.load(an4_audio, sr=16000)
+
+    diarization_pipeline = SortformerDiarization()
+
+    # Simulate streaming
+    chunk_size = 16000  # 1 second
+    for i in range(0, len(signal), chunk_size):
+        chunk = signal[i:i+chunk_size]
+        import asyncio
+        asyncio.run(diarization_pipeline.diarize(chunk))
+
+    for segment in diarization_pipeline.speaker_segments:
+        print(f"Speaker {segment.speaker}: {segment.start:.2f}s - {segment.end:.2f}s")

whisperlivekit/diarization/sortformer_backend_2.py

@@ -0,0 +1,257 @@
+import numpy as np
+import torch
+import logging
+import math
+logger = logging.getLogger(__name__)
+
+try:
+    from nemo.collections.asr.models import SortformerEncLabelModel
+except ImportError:
+    raise SystemExit("""Please use `pip install "git+https://github.com/NVIDIA/NeMo.git@main#egg=nemo_toolkit[asr]"` to use the Sortformer diarization""")
+    
+
+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"))
+    
+# Set the streaming parameters corresponding to 1.04s latency setup. This will affect the streaming feat loader.
+# diar_model.sortformer_modules.chunk_len = 6
+# diar_model.sortformer_modules.spkcache_len = 188
+# diar_model.sortformer_modules.chunk_right_context = 7
+# diar_model.sortformer_modules.fifo_len = 188
+# diar_model.sortformer_modules.spkcache_update_period = 144
+# diar_model.sortformer_modules.log = False
+
+
+# here we change the settings for our goal: speed!
+# we want batches of around 1 second. one frame is 0.08s, so 1s is 12.5 frames. we take 12.
+diar_model.sortformer_modules.chunk_len = 12
+
+# for more speed, we reduce the 'right context'. it's like looking less into the future.
+diar_model.sortformer_modules.chunk_right_context = 1
+
+# we keep the rest same for now
+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()
+
+batch_size = 1
+processed_signal_offset = torch.zeros((batch_size,), dtype=torch.long, device=diar_model.device)
+
+# from nemo.collections.asr.parts.preprocessing.features import FilterbankFeatures
+# from nemo.collections.asr.modules.audio_preprocessing import get_features
+from nemo.collections.asr.modules.audio_preprocessing import AudioToMelSpectrogramPreprocessor
+
+
+def prepare_audio_signal(signal):
+    audio_signal = torch.tensor(signal).unsqueeze(0).to(diar_model.device)
+    audio_signal_length = torch.tensor([audio_signal.shape[1]]).to(diar_model.device)
+    processed_signal, processed_signal_length = AudioToMelSpectrogramPreprocessor(
+            window_size= 0.025, 
+            normalize="NA",
+            n_fft=512,
+            features=128).get_features(audio_signal, audio_signal_length)
+    return processed_signal, processed_signal_length
+
+
+def streaming_feat_loader(
+    feat_seq, feat_seq_length, feat_seq_offset
+):
+    """
+    Load a chunk of feature sequence for streaming inference.
+
+    Args:
+        feat_seq (torch.Tensor): Tensor containing feature sequence
+            Shape: (batch_size, feat_dim, feat frame count)
+        feat_seq_length (torch.Tensor): Tensor containing feature sequence lengths
+            Shape: (batch_size,)
+        feat_seq_offset (torch.Tensor): Tensor containing feature sequence offsets
+            Shape: (batch_size,)
+
+    Returns:
+        chunk_idx (int): Index of the current chunk
+        chunk_feat_seq (torch.Tensor): Tensor containing the chunk of feature sequence
+            Shape: (batch_size, diar frame count, feat_dim)
+        feat_lengths (torch.Tensor): Tensor containing lengths of the chunk of feature sequence
+            Shape: (batch_size,)
+    """
+    feat_len = feat_seq.shape[2]
+    num_chunks = math.ceil(feat_len / (diar_model.sortformer_modules.chunk_len * diar_model.sortformer_modules.subsampling_factor))
+    if False:
+        logging.info(
+            f"feat_len={feat_len}, num_chunks={num_chunks}, "
+            f"feat_seq_length={feat_seq_length}, feat_seq_offset={feat_seq_offset}"
+        )
+
+    stt_feat, end_feat, chunk_idx = 0, 0, 0
+    while end_feat < feat_len:
+        left_offset = min(diar_model.sortformer_modules.chunk_left_context * diar_model.sortformer_modules.subsampling_factor, stt_feat)
+        end_feat = min(stt_feat + diar_model.sortformer_modules.chunk_len * diar_model.sortformer_modules.subsampling_factor, feat_len)
+        right_offset = min(diar_model.sortformer_modules.chunk_right_context * diar_model.sortformer_modules.subsampling_factor, feat_len - end_feat)
+        chunk_feat_seq = feat_seq[:, :, stt_feat - left_offset : end_feat + right_offset]
+        feat_lengths = (feat_seq_length + feat_seq_offset - stt_feat + left_offset).clamp(
+            0, chunk_feat_seq.shape[2]
+        )
+        feat_lengths = feat_lengths * (feat_seq_offset < end_feat)
+        stt_feat = end_feat
+        chunk_feat_seq_t = torch.transpose(chunk_feat_seq, 1, 2)
+        if False:
+            logging.info(
+                f"chunk_idx: {chunk_idx}, "
+                f"chunk_feat_seq_t shape: {chunk_feat_seq_t.shape}, "
+                f"chunk_feat_lengths: {feat_lengths}"
+            )
+        yield chunk_idx, chunk_feat_seq_t, feat_lengths, left_offset, right_offset
+        chunk_idx += 1
+
+
+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(signal, chunks):
+    
+    audio_signal = torch.tensor(signal).unsqueeze(0).to(diar_model.device)
+    audio_signal_length = torch.tensor([audio_signal.shape[1]]).to(diar_model.device)
+    processed_signal, processed_signal_length = AudioToMelSpectrogramPreprocessor(
+            window_size= 0.025, 
+            normalize="NA",
+            n_fft=512,
+            features=128).get_features(audio_signal, audio_signal_length)
+
+    
+    streaming_loader = streaming_feat_loader(processed_signal, processed_signal_length, processed_signal_offset)
+
+    
+    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
+    print(f"Chunk duration: {chunk_duration_seconds} seconds")
+
+    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 streaming_loader:
+        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:]
+                print(chunk_feat_seq_t.shape, total_preds.shape)
+                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
+                    
+        print(l_speakers)
+        """
+        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}]
+        """
+
+if __name__ == '__main__':
+    import librosa
+    an4_audio = 'new_audio_test.mp3'
+    signal, sr = librosa.load(an4_audio,sr=16000) 
+
+    """
+    ground truth:
+    speaker 0 : 0:00 - 0:09
+    speaker 1 : 0:09 - 0:19
+    speaker 2 : 0:19 - 0:25
+    speaker 0 : 0:25 - end
+    """
+
+    # Simulate streaming
+    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(signal, chunks)

whisperlivekit/ffmpeg_manager.py

@@ -0,0 +1,193 @@
+import asyncio
+import logging
+from enum import Enum
+from typing import Optional, Callable
+import contextlib
+
+logger = logging.getLogger(__name__)
+logging.basicConfig(level=logging.INFO)
+
+ERROR_INSTALL_INSTRUCTIONS = """
+FFmpeg is not installed or not found in your system's PATH.
+Please install FFmpeg to enable audio processing.
+
+Installation instructions:
+
+# Ubuntu/Debian:
+sudo apt update && sudo apt install ffmpeg
+
+# macOS (using Homebrew):
+brew install ffmpeg
+
+# Windows:
+# 1. Download the latest static build from https://ffmpeg.org/download.html
+# 2. Extract the archive (e.g., to C:\\FFmpeg).
+# 3. Add the 'bin' directory (e.g., C:\\FFmpeg\\bin) to your system's PATH environment variable.
+
+After installation, please restart the application.
+"""
+
+class FFmpegState(Enum):
+    STOPPED = "stopped"
+    STARTING = "starting"
+    RUNNING = "running"
+    RESTARTING = "restarting"
+    FAILED = "failed"
+
+class FFmpegManager:
+    def __init__(self, sample_rate: int = 16000, channels: int = 1):
+        self.sample_rate = sample_rate
+        self.channels = channels
+
+        self.process: Optional[asyncio.subprocess.Process] = None
+        self._stderr_task: Optional[asyncio.Task] = None
+
+        self.on_error_callback: Optional[Callable[[str], None]] = None
+
+        self.state = FFmpegState.STOPPED
+        self._state_lock = asyncio.Lock()
+
+    async def start(self) -> bool:
+        async with self._state_lock:
+            if self.state != FFmpegState.STOPPED:
+                logger.warning(f"FFmpeg already running in state: {self.state}")
+                return False
+            self.state = FFmpegState.STARTING
+
+        try:
+            cmd = [
+                "ffmpeg",
+                "-hide_banner",
+                "-loglevel", "error",
+                "-i", "pipe:0",
+                "-f", "s16le",
+                "-acodec", "pcm_s16le",
+                "-ac", str(self.channels),
+                "-ar", str(self.sample_rate),
+                "pipe:1"
+            ]
+
+            self.process = await asyncio.create_subprocess_exec(
+                *cmd,
+                stdin=asyncio.subprocess.PIPE,
+                stdout=asyncio.subprocess.PIPE,
+                stderr=asyncio.subprocess.PIPE
+            )
+
+            self._stderr_task = asyncio.create_task(self._drain_stderr())
+
+            async with self._state_lock:
+                self.state = FFmpegState.RUNNING
+
+            logger.info("FFmpeg started.")
+            return True
+
+        except FileNotFoundError:
+            logger.error(ERROR_INSTALL_INSTRUCTIONS)
+            async with self._state_lock:
+                self.state = FFmpegState.FAILED
+            if self.on_error_callback:
+                await self.on_error_callback("ffmpeg_not_found")
+            return False
+
+        except Exception as e:
+            logger.error(f"Error starting FFmpeg: {e}")
+            async with self._state_lock:
+                self.state = FFmpegState.FAILED
+            if self.on_error_callback:
+                await self.on_error_callback("start_failed")
+            return False
+
+    async def stop(self):
+        async with self._state_lock:
+            if self.state == FFmpegState.STOPPED:
+                return
+            self.state = FFmpegState.STOPPED
+
+        if self.process:
+            if self.process.stdin and not self.process.stdin.is_closing():
+                self.process.stdin.close()
+                await self.process.stdin.wait_closed()
+            await self.process.wait()
+            self.process = None
+
+        if self._stderr_task:
+            self._stderr_task.cancel()
+            with contextlib.suppress(asyncio.CancelledError):
+                await self._stderr_task
+
+        logger.info("FFmpeg stopped.")
+
+    async def write_data(self, data: bytes) -> bool:
+        async with self._state_lock:
+            if self.state != FFmpegState.RUNNING:
+                logger.warning(f"Cannot write, FFmpeg state: {self.state}")
+                return False
+
+        try:
+            self.process.stdin.write(data)
+            await self.process.stdin.drain()
+            return True
+        except Exception as e:
+            logger.error(f"Error writing to FFmpeg: {e}")
+            if self.on_error_callback:
+                await self.on_error_callback("write_error")
+            return False
+
+    async def read_data(self, size: int) -> Optional[bytes]:
+        async with self._state_lock:
+            if self.state != FFmpegState.RUNNING:
+                logger.warning(f"Cannot read, FFmpeg state: {self.state}")
+                return None
+
+        try:
+            data = await asyncio.wait_for(
+                self.process.stdout.read(size),
+                timeout=20.0
+            )
+            return data
+        except asyncio.TimeoutError:
+            logger.warning("FFmpeg read timeout.")
+            return None
+        except Exception as e:
+            logger.error(f"Error reading from FFmpeg: {e}")
+            if self.on_error_callback:
+                await self.on_error_callback("read_error")
+            return None
+
+    async def get_state(self) -> FFmpegState:
+        async with self._state_lock:
+            return self.state
+
+    async def restart(self) -> bool:
+        async with self._state_lock:
+            if self.state == FFmpegState.RESTARTING:
+                logger.warning("Restart already in progress.")
+                return False
+            self.state = FFmpegState.RESTARTING
+
+        logger.info("Restarting FFmpeg...")
+
+        try:
+            await self.stop()
+            await asyncio.sleep(1)  # short delay before restarting
+            return await self.start()
+        except Exception as e:
+            logger.error(f"Error during FFmpeg restart: {e}")
+            async with self._state_lock:
+                self.state = FFmpegState.FAILED
+            if self.on_error_callback:
+                await self.on_error_callback("restart_failed")
+            return False
+
+    async def _drain_stderr(self):
+        try:
+            while True:
+                line = await self.process.stderr.readline()
+                if not line:
+                    break
+                logger.debug(f"FFmpeg stderr: {line.decode(errors='ignore').strip()}")
+        except asyncio.CancelledError:
+            logger.info("FFmpeg stderr drain task cancelled.")
+        except Exception as e:
+            logger.error(f"Error draining FFmpeg stderr: {e}")

whisperlivekit/parse_args.py

@@ -58,6 +58,14 @@ def parse_args():
         help="Hugging Face model ID for pyannote.audio embedding model.",
     )
 
+    parser.add_argument(
+        "--diarization-backend",
+        type=str,
+        default="diart",
+        choices=["sortformer", "diart"],
+        help="The diarization backend to use.",
+    )
+
     parser.add_argument(
         "--no-transcription",
         action="store_true",
@@ -74,7 +82,7 @@ def parse_args():
     parser.add_argument(
         "--model",
         type=str,
-        default="tiny",
+        default="small",
         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.",
     )
     
@@ -107,15 +115,15 @@ def parse_args():
     parser.add_argument(
         "--backend",
         type=str,
-        default="faster-whisper",
+        default="simulstreaming",
         choices=["faster-whisper", "whisper_timestamped", "mlx-whisper", "openai-api", "simulstreaming"],
         help="Load only this backend for Whisper processing.",
     )
     parser.add_argument(
-        "--vac",
+        "--no-vac",
         action="store_true",
         default=False,
-        help="Use VAC = voice activity controller. Recommended. Requires torch.",
+        help="Disable VAC = voice activity controller.",
     )
     parser.add_argument(
         "--vac-chunk-size", type=float, default=0.04, help="VAC sample size in seconds."
@@ -242,6 +250,14 @@ def parse_args():
         dest="model_path",
         help="Direct path to the SimulStreaming Whisper .pt model file. Overrides --model for SimulStreaming backend.",
     )
+    
+    simulstreaming_group.add_argument(
+        "--preloaded_model_count",
+        type=int,
+        default=1,
+        dest="preloaded_model_count",
+        help="Optional. Number of models to preload in memory to speed up loading (set up to the expected number of concurrent instances).",
+    )
 
     args = parser.parse_args()
     

whisperlivekit/remove_silences.py

@@ -0,0 +1,110 @@
+from whisperlivekit.timed_objects import ASRToken
+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.end - silence_token.start >= 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, buffer_transcription, buffer_diarization, current_time, vac_detected_silence):
+    if not tokens:
+        return [], buffer_transcription, buffer_diarization
+    last_token = tokens[-1]
+    if tokens and (
+        current_time - last_token.end >= END_SILENCE_DURATION 
+        or 
+        (current_time - last_token.end >= 3 and vac_detected_silence)
+        ):
+        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
+                )
+            )
+        buffer_transcription = "" # for whisperstreaming backend, we should probably validate the buffer has because of the silence
+        buffer_diarization  = ""
+    return tokens, buffer_transcription, buffer_diarization
+    
+
+def handle_silences(tokens, buffer_transcription, buffer_diarization, current_time, vac_detected_silence):
+    tokens = blank_to_silence(tokens) #useful for simulstreaming backend which tends to generate [BLANK_AUDIO] text
+    tokens = no_token_to_silence(tokens)
+    tokens, buffer_transcription, buffer_diarization = ends_with_silence(tokens, buffer_transcription, buffer_diarization, current_time, vac_detected_silence)
+    return tokens, buffer_transcription, buffer_diarization
+     

whisperlivekit/simul_whisper/__init__.py

@@ -0,0 +1,6 @@
+from .backend import SimulStreamingASR, SimulStreamingOnlineProcessor
+
+__all__ = [
+    "SimulStreamingASR",
+    "SimulStreamingOnlineProcessor",
+]

whisperlivekit/simul_whisper/backend.py

@@ -0,0 +1,315 @@
+import sys
+import numpy as np
+import logging
+from typing import List, Tuple, Optional
+import logging
+from whisperlivekit.timed_objects import ASRToken, Transcript
+from whisperlivekit.warmup import load_file
+from whisperlivekit.simul_whisper.license_simulstreaming import SIMULSTREAMING_LICENSE
+from .whisper import load_model, tokenizer
+from .whisper.audio import TOKENS_PER_SECOND
+
+import os
+import gc
+logger = logging.getLogger(__name__)
+
+try:
+    import torch
+    from whisperlivekit.simul_whisper.config import AlignAttConfig
+    from whisperlivekit.simul_whisper.simul_whisper import PaddedAlignAttWhisper
+    from whisperlivekit.simul_whisper.whisper import tokenizer
+except ImportError as e:
+    raise ImportError(
+        """SimulStreaming dependencies are not available.
+        Please install WhisperLiveKit using pip install "whisperlivekit[simulstreaming]".""")
+
+# TOO_MANY_REPETITIONS = 3
+
+class SimulStreamingOnlineProcessor:
+    SAMPLING_RATE = 16000
+
+    def __init__(
+        self,
+        asr,
+        logfile=sys.stderr,
+        warmup_file=None
+    ):        
+        self.asr = asr
+        self.logfile = logfile
+        self.end = 0.0
+        self.global_time_offset = 0.0
+        
+        self.committed: List[ASRToken] = []
+        self.last_result_tokens: List[ASRToken] = []
+        self.load_new_backend()
+        if asr.tokenizer:
+            self.model.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)
+
+    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.
+            self.model.refresh_segment(complete=True)
+            self.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)
+
+    def get_buffer(self):
+        return Transcript(
+            start=None, 
+            end=None, 
+            text='', 
+            probability=None
+        )
+
+    def timestamped_text(self, tokens, generation):
+        """
+        generate timestamped text from tokens and generation data.
+        
+        args:
+            tokens: List of tokens to process
+            generation: Dictionary containing generation progress and optionally results
+            
+        returns:
+            List of tuples containing (start_time, end_time, word) for each word
+        """
+        FRAME_DURATION = 0.02    
+        if "result" in generation:
+            split_words = generation["result"]["split_words"]
+            split_tokens = generation["result"]["split_tokens"]
+        else:
+            split_words, split_tokens = self.model.tokenizer.split_to_word_tokens(tokens)
+        progress = generation["progress"]
+        frames = [p["most_attended_frames"][0] for p in progress]
+        absolute_timestamps = [p["absolute_timestamps"][0] for p in progress]
+        tokens_queue = tokens.copy()
+        timestamped_words = []
+        
+        for word, word_tokens in zip(split_words, split_tokens):
+            # start_frame = None
+            # end_frame = None
+            for expected_token in word_tokens:
+                if not tokens_queue or not frames:
+                    raise ValueError(f"Insufficient tokens or frames for word '{word}'")
+                    
+                actual_token = tokens_queue.pop(0)
+                current_frame = frames.pop(0)
+                current_timestamp = absolute_timestamps.pop(0)
+                if actual_token != expected_token:
+                    raise ValueError(
+                        f"Token mismatch: expected '{expected_token}', "
+                        f"got '{actual_token}' at frame {current_frame}"
+                    )
+                # if start_frame is None:
+                #     start_frame = current_frame
+                # end_frame = current_frame
+            # start_time = start_frame * FRAME_DURATION
+            # end_time = end_frame * FRAME_DURATION
+            start_time = current_timestamp
+            end_time = current_timestamp + 0.1
+            timestamp_entry = (start_time, end_time, word)
+            timestamped_words.append(timestamp_entry)
+            logger.debug(f"TS-WORD:\t{start_time:.2f}\t{end_time:.2f}\t{word}")
+        return timestamped_words
+
+    def process_iter(self, is_last=False) -> Tuple[List[ASRToken], float]:
+        """
+        Process accumulated audio chunks using SimulStreaming.
+        
+        Returns a tuple: (list of committed ASRToken objects, float representing the audio processed up to time).
+        """
+        try:
+            tokens, generation_progress = self.model.infer(is_last=is_last)
+            ts_words = self.timestamped_text(tokens, generation_progress)
+            
+            new_tokens = []
+            for ts_word in ts_words:
+                
+                start, end, word = ts_word
+                token = ASRToken(
+                    start=start,
+                    end=end,
+                    text=word,
+                    probability=0.95  # fake prob. Maybe we can extract it from the model?
+                ).with_offset(
+                    self.global_time_offset
+                )
+                new_tokens.append(token)
+                
+            # identical_tokens = 0
+            # n_new_tokens = len(new_tokens)
+            # if n_new_tokens:
+            
+            self.committed.extend(new_tokens)
+            
+            # if token in self.committed:
+            #     pos = len(self.committed) - 1 - self.committed[::-1].index(token)
+            # if pos:
+            #     for i in range(len(self.committed) - n_new_tokens, -1, -n_new_tokens):
+            #         commited_segment = self.committed[i:i+n_new_tokens]
+            #         if commited_segment == new_tokens:
+            #             identical_segments +=1
+            #             if identical_tokens >= TOO_MANY_REPETITIONS:
+            #                 logger.warning('Too many repetition, model is stuck. Load a new one')
+            #                 self.committed = self.committed[:i]
+            #                 self.load_new_backend()
+            #                 return [], self.end
+
+            # pos = self.committed.rindex(token)
+
+            
+            
+            return new_tokens, self.end
+
+            
+        except Exception as e:
+            logger.exception(f"SimulStreaming processing error: {e}")
+            return [], self.end
+
+    def warmup(self, audio, init_prompt=""):
+        """Warmup the SimulStreaming model."""
+        try:
+            self.model.insert_audio(audio)
+            self.model.infer(True)
+            self.model.refresh_segment(complete=True)
+            logger.info("SimulStreaming model warmed up successfully")
+        except Exception as e:
+            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()
+
+class SimulStreamingASR():
+    """SimulStreaming backend with AlignAtt policy."""
+    sep = ""
+
+    def __init__(self, lan, modelsize=None, cache_dir=None, model_dir=None, logfile=sys.stderr, **kwargs):
+        logger.warning(SIMULSTREAMING_LICENSE)
+        self.logfile = logfile
+        self.transcribe_kargs = {}
+        self.original_language = None if lan == "auto" else lan
+        
+        self.model_path = kwargs.get('model_path', './large-v3.pt')
+        self.frame_threshold = kwargs.get('frame_threshold', 25)
+        self.audio_max_len = kwargs.get('audio_max_len', 20.0)
+        self.audio_min_len = kwargs.get('audio_min_len', 0.0)
+        self.segment_length = kwargs.get('segment_length', 0.5)
+        self.beams = kwargs.get('beams', 1)
+        self.decoder_type = kwargs.get('decoder_type', 'greedy' if self.beams == 1 else 'beam')
+        self.task = kwargs.get('task', 'transcribe')
+        self.cif_ckpt_path = kwargs.get('cif_ckpt_path', None)
+        self.never_fire = kwargs.get('never_fire', False)
+        self.init_prompt = kwargs.get('init_prompt', None)
+        self.static_init_prompt = kwargs.get('static_init_prompt', None)
+        self.max_context_tokens = kwargs.get('max_context_tokens', None)
+        self.warmup_file = kwargs.get('warmup_file', None)
+        self.preload_model_count = kwargs.get('preload_model_count', 1)
+        
+        if model_dir is not None:
+            self.model_path = model_dir
+        elif modelsize 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_path = model_mapping.get(modelsize, f'./{modelsize}.pt')
+        
+        # Set up tokenizer for translation if needed
+        if self.task == "translate":
+            self.tokenizer = self.set_translate_task()
+        else:
+            self.tokenizer = None
+        self.cfg = AlignAttConfig(
+                model_path=self.model_path,
+                segment_length=self.segment_length,
+                frame_threshold=self.frame_threshold,
+                language=self.original_language,
+                audio_max_len=self.audio_max_len,
+                audio_min_len=self.audio_min_len,
+                cif_ckpt_path=self.cif_ckpt_path,
+                decoder_type="beam",
+                beam_size=self.beams,
+                task=self.task,
+                never_fire=self.never_fire,
+                init_prompt=self.init_prompt,
+                max_context_tokens=self.max_context_tokens,
+                static_init_prompt=self.static_init_prompt,
+        )  
+        
+        self.model_name = os.path.basename(self.cfg.model_path).replace(".pt", "")
+        self.model_path = os.path.dirname(os.path.abspath(self.cfg.model_path))
+        self.models = [self.load_model() for i in range(self.preload_model_count)]
+    
+
+
+
+    def load_model(self):
+        whisper_model = load_model(name=self.model_name, download_root=self.model_path)
+        warmup_audio = load_file(self.warmup_file)
+        whisper_model.transcribe(warmup_audio, language=self.original_language)
+        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."""
+        return tokenizer.get_tokenizer(
+            multilingual=True,
+            language=self.model.cfg.language,
+            num_languages=self.model.model.num_languages,
+            task="translate"
+        )
+
+    def transcribe(self, audio):
+        """
+        Warmup is done directly in load_model
+        """
+        pass

whisperlivekit/simul_whisper/beam.py

@@ -0,0 +1,17 @@
+from .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
+
+    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)

whisperlivekit/simul_whisper/config.py

@@ -0,0 +1,29 @@
+# This code was originally in simul_whisper/transcriber/simul_whisper.py . It is adapted a lot for SimulStreaming.
+
+from dataclasses import dataclass, field
+from typing import Literal
+
+@dataclass
+class SimulWhisperConfig:
+    '''Options that are common for all simul policies that could be implemented in SimulWhisper.'''
+    model_path: str
+    language: str = field(default="zh")
+    nonspeech_prob: float = 0.5
+    audio_min_len: float = 1.0
+    decoder_type: Literal["greedy","beam"] = "greedy"
+    beam_size: int = 5
+    task: Literal["transcribe","translate"] = "transcribe"
+    init_prompt: str = field(default=None)
+    static_init_prompt: str = field(default=None)
+    max_context_tokens: int = field(default=None)
+
+@dataclass
+class AlignAttConfig(SimulWhisperConfig):
+    '''Options specific to the AlignAtt policy.'''
+    eval_data_path: str = "tmp"
+    segment_length: float = field(default=1.0, metadata = {"help": "in second"})
+    frame_threshold: int = 4
+    rewind_threshold: int = 200
+    audio_max_len: float = 20.0
+    cif_ckpt_path: str = ""
+    never_fire: bool = False

whisperlivekit/simul_whisper/dual_license_simulstreaming.md

@@ -1,27 +0,0 @@
-
-
-📄 SimulStreaming (https://github.com/ufal/SimulStreaming) Licence
-
-SimulStreaming is dual-licensed:
-
-🔹 Non-Commercial Use
-
-You may use SimulStreaming under the **PolyForm Noncommercial License 1.0.0** if you
-obtain the code through the GitHub repository.  This license is **free of charge**
-and comes with **no obligations** for non-commercial users.
-
-🔸 Commercial Use
-
-Understanding who uses SimulStreaming commercially helps us improve and
-prioritize development. Therefore, we want to **require registration** of those who acquire a commercial licence.
-
-We plan to make the commercial licenceses **affordable** to SMEs and individuals. We
-are considering to provide commercial licenses either for free or for symbolic
-one-time fee, and maybe also provide additional support. You can share your preference via the [questionnaire](https://forms.cloud.microsoft/e/7tCxb4gJfB).
-
-You can also leave your contact [there](https://forms.cloud.microsoft/e/7tCxb4gJfB) to be notified when the commercial licenses become
-available.
-
-✉️ Contact
-
-[Dominik Macháček](https://ufal.mff.cuni.cz/dominik-machacek/), machacek@ufal.mff.cuni.cz

whisperlivekit/simul_whisper/eow_detection.py

@@ -0,0 +1,65 @@
+import torch
+
+# code for the end-of-word detection based on the CIF model proposed in Simul-Whisper
+
+def load_cif(cfg, n_audio_state, device):
+    """cfg: AlignAttConfig, n_audio_state: int, device: torch.device"""
+    cif_linear = torch.nn.Linear(n_audio_state, 1)
+    if cfg.cif_ckpt_path is None or not cfg.cif_ckpt_path:
+        if cfg.never_fire:
+            never_fire = True
+            always_fire = False
+        else:
+            always_fire = True
+            never_fire = False
+    else:
+        always_fire = False
+        never_fire = cfg.never_fire
+        checkpoint = torch.load(cfg.cif_ckpt_path)
+        cif_linear.load_state_dict(checkpoint)
+    cif_linear.to(device)
+    return cif_linear, always_fire, never_fire
+
+
+# from https://github.com/dqqcasia/mosst/blob/master/fairseq/models/speech_to_text/convtransformer_wav2vec_cif.py
+def resize(alphas, target_lengths, threshold=0.999):
+    """
+    alpha in thresh=1.0 | (0.0, +0.21)
+    target_lengths: if None, apply round and resize, else apply scaling
+    """
+    # sum
+    _num = alphas.sum(-1)
+    num = target_lengths.float()
+    # scaling
+    _alphas = alphas * (num / _num)[:, None].repeat(1, alphas.size(1))
+    # rm attention value that exceeds threashold
+    count = 0
+    while len(torch.where(_alphas > threshold)[0]):
+        count += 1
+        if count > 10:
+            break
+        xs, ys = torch.where(_alphas > threshold)
+        for x, y in zip(xs, ys):
+            if _alphas[x][y] >= threshold:
+                mask = _alphas[x].ne(0).float()
+                mean = 0.5 * _alphas[x].sum() / mask.sum()
+                _alphas[x] = _alphas[x] * 0.5 + mean * mask
+
+    return _alphas, _num
+ 
+def fire_at_boundary(chunked_encoder_feature: torch.Tensor, cif_linear):
+    content_mel_len = chunked_encoder_feature.shape[1] # B, T, D
+    alphas = cif_linear(chunked_encoder_feature).squeeze(dim=2) # B, T
+    alphas = torch.sigmoid(alphas)
+    decode_length = torch.round(alphas.sum(-1)).int()
+    alphas, _ = resize(alphas, decode_length)
+    alphas = alphas.squeeze(0) # (T, )
+    threshold = 0.999
+    integrate = torch.cumsum(alphas[:-1], dim=0) # ignore the peak value at the end of the content chunk
+    exceed_count = integrate[-1] // threshold
+    integrate = integrate - exceed_count*1.0 # minus 1 every time intergrate exceed the threshold
+    important_positions = (integrate >= 0).nonzero(as_tuple=True)[0]
+    if important_positions.numel() == 0:
+        return False
+    else:
+        return important_positions[0] >= content_mel_len-2

whisperlivekit/simul_whisper/generation_progress.py

@@ -0,0 +1,43 @@
+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/license_simulstreaming.py

@@ -0,0 +1,5 @@
+SIMULSTREAMING_LICENSE = f"""
+SimulStreaming backend is dual-licensed:
+• Non-Commercial Use: PolyForm Noncommercial License 1.0.0.
+• Commercial Use: Check SimulStreaming README (github.com/ufal/SimulStreaming) for more details.
+"""

whisperlivekit/simul_whisper/simul_whisper.py

@@ -0,0 +1,621 @@
+# This code was originally in simul_whisper/transcriber/simul_whisper.py . It is adapted a lot for SimulStreaming.
+
+import os
+import logging
+
+import torch
+import torch.nn.functional as F
+
+from .whisper import load_model, DecodingOptions, tokenizer
+from .config import AlignAttConfig
+from .whisper.audio import log_mel_spectrogram, TOKENS_PER_SECOND, pad_or_trim, N_SAMPLES, N_FRAMES
+from .whisper.timing import median_filter
+from .whisper.decoding import GreedyDecoder, BeamSearchDecoder, SuppressTokens, detect_language
+from .beam import BeamPyTorchInference
+from .eow_detection import fire_at_boundary, load_cif
+import os
+
+from .token_buffer import TokenBuffer
+
+import numpy as np
+from .generation_progress import *
+
+DEC_PAD = 50257
+logger = logging.getLogger(__name__)
+
+import sys
+import wave
+
+# New features added to the original version of Simul-Whisper: 
+# - large-v3 model support
+# - translation support
+# - beam search
+# - prompt -- static vs. non-static
+# - context
+class PaddedAlignAttWhisper:
+    def __init__(self, cfg: AlignAttConfig, loaded_model=None) -> None:
+        self.log_segments = 0
+        model_name = os.path.basename(cfg.model_path).replace(".pt", "")
+        model_path = os.path.dirname(os.path.abspath(cfg.model_path))
+        if loaded_model:
+            self.model = loaded_model
+        else:
+            self.model = load_model(name=model_name, download_root=model_path)
+
+        logger.info(f"Model dimensions: {self.model.dims}")
+
+        self.decode_options = DecodingOptions(
+            language = cfg.language, 
+            without_timestamps = True,
+            task=cfg.task
+        )
+        self.tokenizer_is_multilingual = not model_name.endswith(".en")
+        self.create_tokenizer(cfg.language if cfg.language != "auto" else None)
+        self.detected_language = cfg.language if cfg.language != "auto" else None
+        
+        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
+
+        if self.cfg.max_context_tokens is None:
+            self.max_context_tokens = self.max_text_len
+        else:
+            self.max_context_tokens = self.cfg.max_context_tokens
+        self.init_context()
+
+        # decoder type: greedy or beam
+        if cfg.decoder_type == "greedy":
+            logger.info("Using greedy decoder")
+            self.token_decoder = GreedyDecoder(0.0, self.tokenizer.eot)
+            self.decoder_type = "greedy"
+
+        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)
+            
+    def remove_hooks(self):
+        print('remove hook')
+        for hook in self.l_hooks:
+            hook.remove()
+
+    def create_tokenizer(self, language=None):
+        self.tokenizer = tokenizer.get_tokenizer(
+            multilingual=self.tokenizer_is_multilingual,  
+            language=language,
+            num_languages=self.model.num_languages,
+            task=self.decode_options.task
+        )
+
+    def init_context(self):
+        kw = {'tokenizer': self.tokenizer, 
+              'device': self.model.device, 
+              'prefix_token_ids': [self.tokenizer.sot_prev]}
+        self.context = TokenBuffer.empty(**kw)
+        if self.cfg.static_init_prompt is not None:
+            self.context = TokenBuffer.from_text(self.cfg.static_init_prompt, **kw)
+        if self.cfg.init_prompt is not None:
+            self.context.text += self.cfg.init_prompt
+
+    def init_tokens(self):
+        logger.debug(f"init tokens, {len(self.segments)}")
+        # init tokens (mandatory prompt)
+        self.initial_tokens = torch.tensor(
+            self.tokenizer.sot_sequence_including_notimestamps, 
+            dtype=torch.long, 
+            device=self.model.device).unsqueeze(0)
+        self.initial_token_length = self.initial_tokens.shape[1]
+        self.sot_index = self.tokenizer.sot_sequence.index(self.tokenizer.sot)
+#        self.segments = []
+        logger.debug(f"init tokens after, {len(self.segments)}")
+        self.tokens = [self.initial_tokens]
+
+    def trim_context(self):
+        logger.info("Trimming context")
+        c = len(self.context.as_token_ids()) - len(self.context.prefix_token_ids)
+#        logger.debug(f"c= {len(self.context.as_token_ids())}, {len(self.context.prefix_token_ids)}")
+        logger.info(f"Context text: {self.context.as_text()}")
+#        logger.debug(f"Context tensor: {self.context.as_tensor()}")
+        l = sum(t.shape[1] for t in self.tokens) + c
+#        logger.debug(f"len {l}, c {c}, max_context_tokens {self.max_context_tokens}")
+        if self.cfg.static_init_prompt is None:
+            after = 0
+        else:
+            after = len(self.cfg.static_init_prompt)
+#        logger.debug(f"len {l}, c {c}, max_context_tokens {self.max_context_tokens}")
+        while c > self.max_context_tokens or l > self.max_text_len - 20:
+            t = self.context.trim_words(after=after)
+            l -= t
+            c -= t
+            logger.debug(f"len {l}, c {c}, max_context_tokens {self.max_context_tokens}")
+            if t == 0:
+                break
+#        logger.debug(f"len {l}, c {c}, max_context_tokens {self.max_context_tokens}")
+        logger.info(f"Context after trim: {self.context.text} (len: {l})")
+
+
+    def logits(self, tokens: torch.Tensor, audio_features: torch.Tensor) -> torch.Tensor:
+        if self.cfg.decoder_type == "greedy":
+            logit = self.model.decoder(tokens, audio_features, kv_cache=self.kv_cache)
+        else:
+            logger.debug(f"Logits shape: {tokens.shape}")
+            logit = self.inference.logits(tokens, audio_features)
+        return logit
+    
+
+    def refresh_segment(self, complete=False):
+
+        logger.debug("Refreshing segment:")
+        self.init_tokens()
+        self.last_attend_frame = -self.cfg.rewind_threshold       
+        self.detected_language = None
+        self.cumulative_time_offset = 0.0
+        self.init_context()
+        logger.debug(f"Context: {self.context}")
+        if not complete and len(self.segments) > 2:
+            logger.debug("keeping last two segments because they are and it is not complete.")
+            self.segments = self.segments[-2:]
+        else:
+            logger.debug("removing all segments.")
+            self.segments = []
+        self.log_segments += 1
+
+
+    def fire_at_boundary(self, chunked_encoder_feature: torch.Tensor):
+        if self.always_fire: return True
+        if self.never_fire: return False
+        return fire_at_boundary(chunked_encoder_feature, self.CIFLinear)
+
+
+    def _current_tokens(self):
+
+        toks = self.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)
+
+        if not self.context.is_empty():
+            context_toks = self.context.as_tensor_beam(self.cfg.beam_size, device=self.model.device)
+            toks = [context_toks] + toks
+
+        # make it one tensor
+        if len(toks) > 1:
+            current_tokens = torch.cat(toks, dim=1)
+        else:
+            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):
+        for i in range(self.cfg.beam_size):
+            logger.debug(self.tokenizer.decode_with_timestamps(tokens[i].tolist()))
+
+    ### audio buffer 
+
+    def segments_len(self):
+        segments_len = sum(s.shape[0] for s in self.segments) / 16000
+        return segments_len
+
+    def _apply_minseglen(self):
+        segments_len = self.segments_len()
+        # wait for long enough audio to start
+        if segments_len < self.cfg.audio_min_len: 
+            logger.debug("waiting for next segment")
+            return False
+        return True
+
+    def insert_audio(self, segment=None):
+        if segment is not None:
+            self.segments.append(segment)
+
+        removed_len = 0
+        # len of audio is bigger than buffer_len. Going to remove the first segment
+        segments_len = self.segments_len()
+        while len(self.segments) > 1 and segments_len > self.cfg.audio_max_len:
+            removed_len = self.segments[0].shape[0] / 16000
+            segments_len -= removed_len
+            self.last_attend_frame -= int(TOKENS_PER_SECOND*removed_len)
+            self.cumulative_time_offset += removed_len  # Track cumulative time removed
+            self.segments = self.segments[1:]
+            logger.debug(f"remove segments: {len(self.segments)} {len(self.tokens)}, cumulative offset: {self.cumulative_time_offset:.2f}s")
+            if len(self.tokens) > 1:
+                self.context.append_token_ids(self.tokens[1][0,:])
+                self.tokens = [self.initial_tokens] + self.tokens[2:]
+        return removed_len
+
+    def _clean_cache(self):
+        '''clean the cache that stores the attention matrices and kv_cache.
+        It must be called every time after generation with the model.'''
+        # cleaning cache
+        self.dec_attns = []
+        self.kv_cache = {}
+        if self.decoder_type == "beam":
+            self.inference.kv_cache = self.kv_cache
+            self.token_decoder.reset()
+
+    @torch.no_grad()
+    def lang_id(self, encoder_features):
+        """Language detection from encoder features.
+        This code is trimmed and copy-pasted from whisper.decoding.detect_language .
+        """
+    
+        # forward pass using a single token, startoftranscript
+        n_audio = encoder_features.shape[0]
+        x = torch.tensor([[self.tokenizer.sot]] * n_audio).to(self.model.device)  # [n_audio, 1]
+        logits = self.model.logits(x, encoder_features)[:, 0]
+
+        # collect detected languages; suppress all non-language tokens
+        mask = torch.ones(logits.shape[-1], dtype=torch.bool)
+        mask[list(self.tokenizer.all_language_tokens)] = False
+        logits[:, mask] = -np.inf
+        language_tokens = logits.argmax(dim=-1)
+        language_token_probs = logits.softmax(dim=-1).cpu()
+        language_probs = [
+            {
+                c: language_token_probs[i, j].item()
+                for j, c in zip(self.tokenizer.all_language_tokens, self.tokenizer.all_language_codes)
+            }
+            for i in range(n_audio)
+        ]
+
+        single = encoder_features.ndim == 2
+        if single:
+            language_tokens = language_tokens[0]
+            language_probs = language_probs[0]
+
+        self._clean_cache()
+        return language_tokens, language_probs
+
+    ### transcription / translation
+
+    @torch.no_grad()
+    def infer(self, is_last=False):
+        new_segment = True
+        if len(self.segments) == 0:
+            logger.debug("No segments, nothing to do")
+            return [], {}
+        if not self._apply_minseglen():
+            logger.debug(f"applied minseglen {self.cfg.audio_min_len} > {self.segments_len()}.")
+            input_segments = torch.cat(self.segments, dim=0)
+            return [], {}
+
+        # input_segments is concatenation of audio, it's one array
+        if len(self.segments) > 1:
+            input_segments = torch.cat(self.segments, dim=0)
+        else:
+            input_segments = self.segments[0]
+
+
+        
+        # mel + padding to 30s
+        mel_padded = log_mel_spectrogram(input_segments, n_mels=self.model.dims.n_mels, padding=N_SAMPLES, 
+                                            device=self.model.device).unsqueeze(0)
+        # trim to 3000
+        mel = pad_or_trim(mel_padded, N_FRAMES)
+
+        # the len of actual audio
+        content_mel_len = int((mel_padded.shape[2] - mel.shape[2])/2)
+
+        # encode
+        encoder_feature = self.model.encoder(mel)
+
+#        logger.debug(f"Encoder feature shape: {encoder_feature.shape}")
+#        if mel.shape[-2:] != (self.model.dims.n_audio_ctx, self.model.dims.n_audio_state):
+#            logger.debug("mel ")
+        if self.cfg.language == "auto" and self.detected_language is None:
+            language_tokens, language_probs = self.lang_id(encoder_feature) 
+            logger.debug(f"Language tokens: {language_tokens}, probs: {language_probs}")
+            top_lan, p = max(language_probs[0].items(), key=lambda x: x[1])
+            logger.info(f"Detected language: {top_lan} with p={p:.4f}")
+            #self.tokenizer.language = top_lan
+            #self.tokenizer.__post_init__()
+            self.create_tokenizer(top_lan)
+            self.detected_language = top_lan
+            self.init_tokens()
+            logger.info(f"Tokenizer language: {self.tokenizer.language}, {self.tokenizer.sot_sequence_including_notimestamps}")
+
+        self.trim_context()
+        current_tokens = self._current_tokens()
+#        
+        fire_detected = self.fire_at_boundary(encoder_feature[:, :content_mel_len, :])
+
+
+        ####################### Decoding loop
+        logger.info("Decoding loop starts\n")
+
+        sum_logprobs = torch.zeros(self.cfg.beam_size, device=mel.device)
+        completed = False
+
+        attn_of_alignment_heads = None
+        most_attended_frame = None
+
+        token_len_before_decoding = current_tokens.shape[1]
+        
+        generation_progress = []
+        generation = {
+            "starting_tokens": BeamTokens(current_tokens[0,:].clone(), self.cfg.beam_size),
+            "token_len_before_decoding": token_len_before_decoding,
+            #"fire_detected": fire_detected,
+            "frames_len": content_mel_len,
+            "frames_threshold": 4 if is_last else self.cfg.frame_threshold,
+
+            # to be filled later
+            "logits_starting": None,
+
+            # to be filled later
+            "no_speech_prob": None,
+            "no_speech": False,
+
+            # to be filled in the loop
+            "progress": generation_progress,
+        }
+        while not completed and current_tokens.shape[1] < self.max_text_len: # bos is 3 tokens
+            generation_progress_loop = []
+
+            if new_segment:
+                tokens_for_logits = current_tokens
+            else:
+                # only need to use the last token except in the first forward pass
+                tokens_for_logits = current_tokens[:,-1:]
+
+            logits = self.logits(tokens_for_logits, encoder_feature) # B, len(tokens), token dict size
+            if new_segment:
+                generation["logits_starting"] = Logits(logits[:,:,:])
+
+            if new_segment and self.tokenizer.no_speech is not None:
+                probs_at_sot = logits[:, self.sot_index, :].float().softmax(dim=-1)
+                no_speech_probs = probs_at_sot[:, self.tokenizer.no_speech].tolist()
+                generation["no_speech_prob"] = no_speech_probs[0]
+                if no_speech_probs[0] > self.cfg.nonspeech_prob:
+                    generation["no_speech"] = True
+                    logger.info("no speech, stop")
+                    break
+
+            logits = logits[:, -1, :] # logits for the last token
+            generation_progress_loop.append(("logits_before_suppress",Logits(logits)))
+
+            # supress blank tokens only at the beginning of the segment
+            if new_segment:
+                logits[:, self.tokenizer.encode(" ") + [self.tokenizer.eot]] = -np.inf
+            new_segment = False
+            self.suppress_tokens(logits)
+            #generation_progress_loop.append(("logits_after_suppres",BeamLogits(logits[0,:].clone(), self.cfg.beam_size)))
+            generation_progress_loop.append(("logits_after_suppress",Logits(logits)))
+
+            current_tokens, completed = self.token_decoder.update(current_tokens, logits, sum_logprobs)
+            generation_progress_loop.append(("beam_tokens",Tokens(current_tokens[:,-1].clone())))
+            generation_progress_loop.append(("sum_logprobs",sum_logprobs.tolist()))
+            generation_progress_loop.append(("completed",completed))
+
+            logger.debug(f"Decoding completed: {completed}, sum_logprobs: {sum_logprobs.tolist()}, tokens: ")
+            self.debug_print_tokens(current_tokens)
+
+
+            # if self.decoder_type == "beam":
+            #     logger.debug(f"Finished sequences: {self.token_decoder.finished_sequences}")
+
+            #     logprobs = F.log_softmax(logits.float(), dim=-1)
+            #     idx = 0
+            #     logger.debug(f"Beam search topk: {logprobs[idx].topk(self.cfg.beam_size + 1)}")
+            #     logger.debug(f"Greedy search argmax: {logits.argmax(dim=-1)}")
+            # if completed:
+            #     self.debug_print_tokens(current_tokens)
+
+            #     logger.debug("decode stopped because decoder completed")
+
+            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)
+#            logger.debug(str(attn_of_alignment_heads.shape) + " tttady")
+            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)
+#            logger.debug(str(attn_of_alignment_heads.shape) + " po mean")
+            attn_of_alignment_heads = attn_of_alignment_heads[:,:, :content_mel_len]
+#            logger.debug(str(attn_of_alignment_heads.shape) + " pak ")
+
+            # for each beam, the most attended frame is:
+            most_attended_frames = torch.argmax(attn_of_alignment_heads[:,-1,:], dim=-1)
+            generation_progress_loop.append(("most_attended_frames",most_attended_frames.clone().tolist()))
+            
+            # Calculate absolute timestamps accounting for cumulative offset
+            absolute_timestamps = [(frame * 0.02 + self.cumulative_time_offset) for frame in most_attended_frames.tolist()]
+            generation_progress_loop.append(("absolute_timestamps", absolute_timestamps))
+            
+            logger.debug(str(most_attended_frames.tolist()) + " most att frames")
+            logger.debug(f"Absolute timestamps: {absolute_timestamps} (offset: {self.cumulative_time_offset:.2f}s)")
+
+            most_attended_frame = most_attended_frames[0].item()
+
+
+            generation_progress.append(dict(generation_progress_loop))
+            logger.debug("current tokens" + str(current_tokens.shape))
+            if completed:
+            #    # stripping the last token, the eot
+                current_tokens = current_tokens[:, :-1]
+                break
+            
+            # for some rare cases where the attention fails
+            if not is_last and self.last_attend_frame - most_attended_frame > self.cfg.rewind_threshold:
+                # TODO: check this
+                if current_tokens.shape[1] > 1 and current_tokens[0, -2] >= DEC_PAD:
+                    logger.debug("ommit rewinding from special tokens")
+                    self.last_attend_frame = most_attended_frame
+                else:
+                    logger.debug(
+                        f"[rewind detected] current attention pos: {most_attended_frame}, "
+                        f"last attention pos: {self.last_attend_frame}; omit this segment")
+                    self.last_attend_frame = -self.cfg.rewind_threshold
+                    current_tokens = torch.cat(self.tokens, dim=1) if len(self.tokens) > 0 else self.tokens[0]
+                    break
+            else:
+                self.last_attend_frame = most_attended_frame
+
+            if content_mel_len - most_attended_frame <= (4 if is_last else self.cfg.frame_threshold):
+                logger.debug(f"attention reaches the end: {most_attended_frame}/{content_mel_len}")
+                # stripping the last token, the one that is attended too close to the end
+                current_tokens = current_tokens[:, :-1]
+                break
+        
+            # debug print
+            for i in range(self.cfg.beam_size):
+                logger.debug("attn: {}, current pos: {}, current token: {}({})".format(
+                    attn_of_alignment_heads.shape if attn_of_alignment_heads is not None else None,
+                    most_attended_frames[i], 
+                    current_tokens[i, -1].item(),
+                    self.tokenizer.decode([current_tokens[i, -1].item()])
+                ))
+
+#        for k,v in generation.items():
+#            print(k,v,file=sys.stderr)
+#        for x in generation_progress:
+#            for y in x.items():
+#                print("\t\t",*y,file=sys.stderr)
+#            print("\t","----", file=sys.stderr)
+#        print("\t", "end of generation_progress_loop", file=sys.stderr)
+        #    sys.exit(1)
+        ####################### End of decoding loop
+
+        logger.info("End of decoding loop")
+
+        # if attn_of_alignment_heads is not None:
+        #     seg_len = int(segment.shape[0] / 16000 * TOKENS_PER_SECOND)
+
+        #     # Lets' now consider only the top hypothesis in the beam search
+        #     top_beam_attn_of_alignment_heads = attn_of_alignment_heads[0]
+
+        #     # debug print: how is the new token attended?
+        #     new_token_attn = top_beam_attn_of_alignment_heads[token_len_before_decoding:, -seg_len:]
+        #     logger.debug(f"New token attention shape: {new_token_attn.shape}")
+        #     if new_token_attn.shape[0] == 0:  # it's not attended in the current audio segment
+        #         logger.debug("no token generated")
+        #     else:  # it is, and the max attention is:
+        #         new_token_max_attn, _ = new_token_attn.max(dim=-1)
+        #         logger.debug(f"segment max attention: {new_token_max_attn.mean().item()/len(self.segments)}")
+
+
+        # let's now operate only with the top beam hypothesis
+        tokens_to_split = current_tokens[0, token_len_before_decoding:]
+        if fire_detected or is_last:
+            new_hypothesis = tokens_to_split.flatten().tolist()
+        else:
+            # going to truncate the tokens after the last space
+            split_words, split_tokens = self.tokenizer.split_to_word_tokens(tokens_to_split.tolist())
+            generation["result"] = {"split_words": split_words[:-1], "split_tokens": split_tokens[:-1]}
+            generation["result_truncated"] = {"split_words": split_words[-1:], "split_tokens": split_tokens[-1:]}
+
+#            text_to_split = self.tokenizer.decode(tokens_to_split)
+#            logger.debug(f"text_to_split: {text_to_split}")
+#            logger.debug("text at current step: {}".format(text_to_split.replace(" ", "<space>")))
+#            text_before_space = " ".join(text_to_split.split(" ")[:-1])
+#            logger.debug("before the last space: {}".format(text_before_space.replace(" ", "<space>")))
+            if len(split_words) > 1:
+                new_hypothesis = [i for sublist in split_tokens[:-1] for i in sublist]  
+            else:
+                new_hypothesis = []
+
+
+        ### 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.model.device,
+        )
+        self.tokens.append(new_tokens)
+        # TODO: test if this is redundant or not
+#        ret = ret[ret<DEC_PAD]
+
+        logger.info(f"Output: {self.tokenizer.decode(new_hypothesis)}")
+        
+        self._clean_cache()
+
+        return new_hypothesis, generation

whisperlivekit/simul_whisper/token_buffer.py

@@ -54,8 +54,8 @@ class TokenBuffer:
 
         ids = tokenizer.encode(self.text[after:])
         words, wids = self.tokenizer.split_to_word_tokens(ids)
-        print(words, file=sys.stderr)
-        print(wids, file=sys.stderr)
+#        print(words, file=sys.stderr)
+#        print(wids, file=sys.stderr)
         if not words:
             return 0
         self.text = self.text[:after] + "".join(words[num:])

whisperlivekit/simul_whisper/whisper/__init__.py

@@ -0,0 +1,160 @@
+import hashlib
+import io
+import os
+import urllib
+import warnings
+from typing import List, Optional, Union
+
+import torch
+from tqdm import tqdm
+
+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__
+
+_MODELS = {
+    "tiny.en": "https://openaipublic.azureedge.net/main/whisper/models/d3dd57d32accea0b295c96e26691aa14d8822fac7d9d27d5dc00b4ca2826dd03/tiny.en.pt",
+    "tiny": "https://openaipublic.azureedge.net/main/whisper/models/65147644a518d12f04e32d6f3b26facc3f8dd46e5390956a9424a650c0ce22b9/tiny.pt",
+    "base.en": "https://openaipublic.azureedge.net/main/whisper/models/25a8566e1d0c1e2231d1c762132cd20e0f96a85d16145c3a00adf5d1ac670ead/base.en.pt",
+    "base": "https://openaipublic.azureedge.net/main/whisper/models/ed3a0b6b1c0edf879ad9b11b1af5a0e6ab5db9205f891f668f8b0e6c6326e34e/base.pt",
+    "small.en": "https://openaipublic.azureedge.net/main/whisper/models/f953ad0fd29cacd07d5a9eda5624af0f6bcf2258be67c92b79389873d91e0872/small.en.pt",
+    "small": "https://openaipublic.azureedge.net/main/whisper/models/9ecf779972d90ba49c06d968637d720dd632c55bbf19d441fb42bf17a411e794/small.pt",
+    "medium.en": "https://openaipublic.azureedge.net/main/whisper/models/d7440d1dc186f76616474e0ff0b3b6b879abc9d1a4926b7adfa41db2d497ab4f/medium.en.pt",
+    "medium": "https://openaipublic.azureedge.net/main/whisper/models/345ae4da62f9b3d59415adc60127b97c714f32e89e936602e85993674d08dcb1/medium.pt",
+    "large-v1": "https://openaipublic.azureedge.net/main/whisper/models/e4b87e7e0bf463eb8e6956e646f1e277e901512310def2c24bf0e11bd3c28e9a/large-v1.pt",
+    "large-v2": "https://openaipublic.azureedge.net/main/whisper/models/81f7c96c852ee8fc832187b0132e569d6c3065a3252ed18e56effd0b6a73e524/large-v2.pt",
+    "large-v3": "https://openaipublic.azureedge.net/main/whisper/models/e5b1a55b89c1367dacf97e3e19bfd829a01529dbfdeefa8caeb59b3f1b81dadb/large-v3.pt",
+    "large": "https://openaipublic.azureedge.net/main/whisper/models/e5b1a55b89c1367dacf97e3e19bfd829a01529dbfdeefa8caeb59b3f1b81dadb/large-v3.pt",
+    "large-v3-turbo": "https://openaipublic.azureedge.net/main/whisper/models/aff26ae408abcba5fbf8813c21e62b0941638c5f6eebfb145be0c9839262a19a/large-v3-turbo.pt",
+    "turbo": "https://openaipublic.azureedge.net/main/whisper/models/aff26ae408abcba5fbf8813c21e62b0941638c5f6eebfb145be0c9839262a19a/large-v3-turbo.pt",
+}
+
+# base85-encoded (n_layers, n_heads) boolean arrays indicating the cross-attention heads that are
+# highly correlated to the word-level timing, i.e. the alignment between audio and text tokens.
+_ALIGNMENT_HEADS = {
+    "tiny.en": b"ABzY8J1N>@0{>%R00Bk>$p{7v037`oCl~+#00",
+    "tiny": b"ABzY8bu8Lr0{>%RKn9Fp%m@SkK7Kt=7ytkO",
+    "base.en": b"ABzY8;40c<0{>%RzzG;p*o+Vo09|#PsxSZm00",
+    "base": b"ABzY8KQ!870{>%RzyTQH3`Q^yNP!>##QT-<FaQ7m",
+    "small.en": b"ABzY8>?_)10{>%RpeA61k&I|OI3I$65C{;;pbCHh0B{qLQ;+}v00",
+    "small": b"ABzY8DmU6=0{>%Rpa?J`kvJ6qF(V^F86#Xh7JUGMK}P<N0000",
+    "medium.en": b"ABzY8usPae0{>%R7<zz_OvQ{)4kMa0BMw6u5rT}kRKX;$NfYBv00*Hl@qhsU00",
+    "medium": b"ABzY8B0Jh+0{>%R7}kK1fFL7w6%<-Pf*t^=N)Qr&0RR9",
+    "large-v1": b"ABzY8r9j$a0{>%R7#4sLmoOs{s)o3~84-RPdcFk!JR<kSfC2yj",
+    "large-v2": b"ABzY8zd+h!0{>%R7=D0pU<_bnWW*tkYAhobTNnu$jnkEkXqp)j;w1Tzk)UH3X%SZd&fFZ2fC2yj",
+    "large-v3": b"ABzY8gWO1E0{>%R7(9S+Kn!D~%ngiGaR?*L!iJG9p-nab0JQ=-{D1-g00",
+    "large": b"ABzY8gWO1E0{>%R7(9S+Kn!D~%ngiGaR?*L!iJG9p-nab0JQ=-{D1-g00",
+    "large-v3-turbo": b"ABzY8j^C+e0{>%RARaKHP%t(lGR*)0g!tONPyhe`",
+    "turbo": b"ABzY8j^C+e0{>%RARaKHP%t(lGR*)0g!tONPyhe`",
+}
+
+
+def _download(url: str, root: str, in_memory: bool) -> Union[bytes, str]:
+    os.makedirs(root, exist_ok=True)
+
+    expected_sha256 = url.split("/")[-2]
+    download_target = os.path.join(root, os.path.basename(url))
+
+    if os.path.exists(download_target) and not os.path.isfile(download_target):
+        raise RuntimeError(f"{download_target} exists and is not a regular file")
+
+    if os.path.isfile(download_target):
+        with open(download_target, "rb") as f:
+            model_bytes = f.read()
+        if hashlib.sha256(model_bytes).hexdigest() == expected_sha256:
+            return model_bytes if in_memory else download_target
+        else:
+            warnings.warn(
+                f"{download_target} exists, but the SHA256 checksum does not match; re-downloading the file"
+            )
+
+    with urllib.request.urlopen(url) as source, open(download_target, "wb") as output:
+        with tqdm(
+            total=int(source.info().get("Content-Length")),
+            ncols=80,
+            unit="iB",
+            unit_scale=True,
+            unit_divisor=1024,
+        ) as loop:
+            while True:
+                buffer = source.read(8192)
+                if not buffer:
+                    break
+
+                output.write(buffer)
+                loop.update(len(buffer))
+
+    model_bytes = open(download_target, "rb").read()
+    if hashlib.sha256(model_bytes).hexdigest() != expected_sha256:
+        raise RuntimeError(
+            "Model has been downloaded but the SHA256 checksum does not not match. Please retry loading the model."
+        )
+
+    return model_bytes if in_memory else download_target
+
+
+def available_models() -> List[str]:
+    """Returns the names of available models"""
+    return list(_MODELS.keys())
+
+
+def load_model(
+    name: str,
+    device: Optional[Union[str, torch.device]] = None,
+    download_root: str = None,
+    in_memory: bool = False,
+) -> Whisper:
+    """
+    Load a Whisper ASR model
+
+    Parameters
+    ----------
+    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.
+    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
+
+    Returns
+    -------
+    model : Whisper
+        The Whisper ASR model instance
+    """
+
+    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")
+
+    if name in _MODELS:
+        checkpoint_file = _download(_MODELS[name], download_root, in_memory)
+        alignment_heads = _ALIGNMENT_HEADS[name]
+    elif os.path.isfile(name):
+        checkpoint_file = open(name, "rb").read() if in_memory else name
+        alignment_heads = None
+    else:
+        raise RuntimeError(
+            f"Model {name} not found; available models = {available_models()}"
+        )
+
+    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 = ModelDimensions(**checkpoint["dims"])
+    model = Whisper(dims)
+    model.load_state_dict(checkpoint["model_state_dict"])
+
+    if alignment_heads is not None:
+        model.set_alignment_heads(alignment_heads)
+
+    return model.to(device)

whisperlivekit/simul_whisper/whisper/__main__.py

@@ -0,0 +1,3 @@
+from .transcribe import cli
+
+cli()

whisperlivekit/simul_whisper/whisper/audio.py

@@ -0,0 +1,157 @@
+import os
+from functools import lru_cache
+from subprocess import CalledProcessError, run
+from typing import Optional, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+from .utils import exact_div
+
+# hard-coded audio hyperparameters
+SAMPLE_RATE = 16000
+N_FFT = 400
+HOP_LENGTH = 160
+CHUNK_LENGTH = 30
+N_SAMPLES = CHUNK_LENGTH * SAMPLE_RATE  # 480000 samples in a 30-second chunk
+N_FRAMES = exact_div(N_SAMPLES, HOP_LENGTH)  # 3000 frames in a mel spectrogram input
+
+N_SAMPLES_PER_TOKEN = HOP_LENGTH * 2  # the initial convolutions has stride 2
+FRAMES_PER_SECOND = exact_div(SAMPLE_RATE, HOP_LENGTH)  # 10ms per audio frame
+TOKENS_PER_SECOND = exact_div(SAMPLE_RATE, N_SAMPLES_PER_TOKEN)  # 20ms per audio token
+
+
+def load_audio(file: str, sr: int = SAMPLE_RATE):
+    """
+    Open an audio file and read as mono waveform, resampling as necessary
+
+    Parameters
+    ----------
+    file: str
+        The audio file to open
+
+    sr: int
+        The sample rate to resample the audio if necessary
+
+    Returns
+    -------
+    A NumPy array containing the audio waveform, in float32 dtype.
+    """
+
+    # This launches a subprocess to decode audio while down-mixing
+    # and resampling as necessary.  Requires the ffmpeg CLI in PATH.
+    # fmt: off
+    cmd = [
+        "ffmpeg",
+        "-nostdin",
+        "-threads", "0",
+        "-i", file,
+        "-f", "s16le",
+        "-ac", "1",
+        "-acodec", "pcm_s16le",
+        "-ar", str(sr),
+        "-"
+    ]
+    # fmt: on
+    try:
+        out = run(cmd, capture_output=True, check=True).stdout
+    except CalledProcessError as e:
+        raise RuntimeError(f"Failed to load audio: {e.stderr.decode()}") from e
+
+    return np.frombuffer(out, np.int16).flatten().astype(np.float32) / 32768.0
+
+
+def pad_or_trim(array, length: int = N_SAMPLES, *, axis: int = -1):
+    """
+    Pad or trim the audio array to N_SAMPLES, as expected by the encoder.
+    """
+    if torch.is_tensor(array):
+        if array.shape[axis] > length:
+            array = array.index_select(
+                dim=axis, index=torch.arange(length, device=array.device)
+            )
+
+        if array.shape[axis] < length:
+            pad_widths = [(0, 0)] * array.ndim
+            pad_widths[axis] = (0, length - array.shape[axis])
+            array = F.pad(array, [pad for sizes in pad_widths[::-1] for pad in sizes])
+    else:
+        if array.shape[axis] > length:
+            array = array.take(indices=range(length), axis=axis)
+
+        if array.shape[axis] < length:
+            pad_widths = [(0, 0)] * array.ndim
+            pad_widths[axis] = (0, length - array.shape[axis])
+            array = np.pad(array, pad_widths)
+
+    return array
+
+
+@lru_cache(maxsize=None)
+def mel_filters(device, n_mels: int) -> torch.Tensor:
+    """
+    load the mel filterbank matrix for projecting STFT into a Mel spectrogram.
+    Allows decoupling librosa dependency; saved using:
+
+        np.savez_compressed(
+            "mel_filters.npz",
+            mel_80=librosa.filters.mel(sr=16000, n_fft=400, n_mels=80),
+            mel_128=librosa.filters.mel(sr=16000, n_fft=400, n_mels=128),
+        )
+    """
+    assert n_mels in {80, 128}, f"Unsupported n_mels: {n_mels}"
+
+    filters_path = os.path.join(os.path.dirname(__file__), "assets", "mel_filters.npz")
+    with np.load(filters_path, allow_pickle=False) as f:
+        return torch.from_numpy(f[f"mel_{n_mels}"]).to(device)
+
+
+def log_mel_spectrogram(
+    audio: Union[str, np.ndarray, torch.Tensor],
+    n_mels: int = 80,
+    padding: int = 0,
+    device: Optional[Union[str, torch.device]] = None,
+):
+    """
+    Compute the log-Mel spectrogram of
+
+    Parameters
+    ----------
+    audio: Union[str, np.ndarray, torch.Tensor], shape = (*)
+        The path to audio or either a NumPy array or Tensor containing the audio waveform in 16 kHz
+
+    n_mels: int
+        The number of Mel-frequency filters, only 80 and 128 are supported
+
+    padding: int
+        Number of zero samples to pad to the right
+
+    device: Optional[Union[str, torch.device]]
+        If given, the audio tensor is moved to this device before STFT
+
+    Returns
+    -------
+    torch.Tensor, shape = (n_mels, n_frames)
+        A Tensor that contains the Mel spectrogram
+    """
+    if not torch.is_tensor(audio):
+        if isinstance(audio, str):
+            audio = load_audio(audio)
+        audio = torch.from_numpy(audio)
+
+    if device is not None:
+        audio = audio.to(device)
+    if padding > 0:
+        audio = F.pad(audio, (0, padding))
+    window = torch.hann_window(N_FFT).to(audio.device)
+    stft = torch.stft(audio, N_FFT, HOP_LENGTH, window=window, return_complex=True)
+    magnitudes = stft[..., :-1].abs() ** 2
+
+    filters = mel_filters(audio.device, n_mels)
+    mel_spec = filters @ magnitudes
+
+    log_spec = torch.clamp(mel_spec, min=1e-10).log10()
+    log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
+    log_spec = (log_spec + 4.0) / 4.0
+    return log_spec

whisperlivekit/simul_whisper/whisper/decoding.py

@@ -0,0 +1,826 @@
+from dataclasses import dataclass, field, replace
+from typing import TYPE_CHECKING, Dict, Iterable, List, Optional, Sequence, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import Tensor
+from torch.distributions import Categorical
+
+from .audio import CHUNK_LENGTH
+from .tokenizer import Tokenizer, get_tokenizer
+from .utils import compression_ratio
+
+if TYPE_CHECKING:
+    from .model import Whisper
+
+
+@torch.no_grad()
+def detect_language(
+    model: "Whisper", mel: Tensor, tokenizer: Tokenizer = None
+) -> Tuple[Tensor, List[dict]]:
+    """
+    Detect the spoken language in the audio, and return them as list of strings, along with the ids
+    of the most probable language tokens and the probability distribution over all language tokens.
+    This is performed outside the main decode loop in order to not interfere with kv-caching.
+
+    Returns
+    -------
+    language_tokens : Tensor, shape = (n_audio,)
+        ids of the most probable language tokens, which appears after the startoftranscript token.
+    language_probs : List[Dict[str, float]], length = n_audio
+        list of dictionaries containing the probability distribution over all languages.
+    """
+    if tokenizer is None:
+        tokenizer = get_tokenizer(
+            model.is_multilingual, num_languages=model.num_languages
+        )
+    if (
+        tokenizer.language is None
+        or tokenizer.language_token not in tokenizer.sot_sequence
+    ):
+        raise ValueError(
+            "This model doesn't have language tokens so it can't perform lang id"
+        )
+
+    single = mel.ndim == 2
+    if single:
+        mel = mel.unsqueeze(0)
+
+    # skip encoder forward pass if already-encoded audio features were given
+    if mel.shape[-2:] != (model.dims.n_audio_ctx, model.dims.n_audio_state):
+        mel = model.encoder(mel)
+
+    # forward pass using a single token, startoftranscript
+    n_audio = mel.shape[0]
+    x = torch.tensor([[tokenizer.sot]] * n_audio).to(mel.device)  # [n_audio, 1]
+    logits = model.logits(x, mel)[:, 0]
+
+    # collect detected languages; suppress all non-language tokens
+    mask = torch.ones(logits.shape[-1], dtype=torch.bool)
+    mask[list(tokenizer.all_language_tokens)] = False
+    logits[:, mask] = -np.inf
+    language_tokens = logits.argmax(dim=-1)
+    language_token_probs = logits.softmax(dim=-1).cpu()
+    language_probs = [
+        {
+            c: language_token_probs[i, j].item()
+            for j, c in zip(tokenizer.all_language_tokens, tokenizer.all_language_codes)
+        }
+        for i in range(n_audio)
+    ]
+
+    if single:
+        language_tokens = language_tokens[0]
+        language_probs = language_probs[0]
+
+    return language_tokens, language_probs
+
+
+@dataclass(frozen=True)
+class DecodingOptions:
+    # whether to perform X->X "transcribe" or X->English "translate"
+    task: str = "transcribe"
+
+    # language that the audio is in; uses detected language if None
+    language: Optional[str] = None
+
+    # sampling-related options
+    temperature: float = 0.0
+    sample_len: Optional[int] = None  # maximum number of tokens to sample
+    best_of: Optional[int] = None  # number of independent sample trajectories, if t > 0
+    beam_size: Optional[int] = None  # number of beams in beam search, if t == 0
+    patience: Optional[float] = None  # patience in beam search (arxiv:2204.05424)
+
+    # "alpha" in Google NMT, or None for length norm, when ranking generations
+    # to select which to return among the beams or best-of-N samples
+    length_penalty: Optional[float] = None
+
+    # text or tokens to feed as the prompt or the prefix; for more info:
+    # https://github.com/openai/whisper/discussions/117#discussioncomment-3727051
+    prompt: Optional[Union[str, List[int]]] = None  # for the previous context
+    prefix: Optional[Union[str, List[int]]] = None  # to prefix the current context
+
+    # list of tokens ids (or comma-separated token ids) to suppress
+    # "-1" will suppress a set of symbols as defined in `tokenizer.non_speech_tokens()`
+    suppress_tokens: Optional[Union[str, Iterable[int]]] = "-1"
+    suppress_blank: bool = True  # this will suppress blank outputs
+
+    # timestamp sampling options
+    without_timestamps: bool = False  # use <|notimestamps|> to sample text tokens only
+    max_initial_timestamp: Optional[float] = 1.0
+
+    # implementation details
+    fp16: bool = True  # use fp16 for most of the calculation
+
+
+@dataclass(frozen=True)
+class DecodingResult:
+    audio_features: Tensor
+    language: str
+    language_probs: Optional[Dict[str, float]] = None
+    tokens: List[int] = field(default_factory=list)
+    text: str = ""
+    avg_logprob: float = np.nan
+    no_speech_prob: float = np.nan
+    temperature: float = np.nan
+    compression_ratio: float = np.nan
+
+
+class Inference:
+    def logits(self, tokens: Tensor, audio_features: Tensor) -> Tensor:
+        """Perform a forward pass on the decoder and return per-token logits"""
+        raise NotImplementedError
+
+    def rearrange_kv_cache(self, source_indices) -> None:
+        """Update the key-value cache according to the updated beams"""
+        raise NotImplementedError
+
+    def cleanup_caching(self) -> None:
+        """Clean up any resources or hooks after decoding is finished"""
+        pass
+
+
+class PyTorchInference(Inference):
+    def __init__(self, model: "Whisper", initial_token_length: int):
+        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
+
+    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:]
+
+        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()
+
+
+class SequenceRanker:
+    def rank(
+        self, tokens: List[List[Tensor]], sum_logprobs: List[List[float]]
+    ) -> List[int]:
+        """
+        Given a list of groups of samples and their cumulative log probabilities,
+        return the indices of the samples in each group to select as the final result
+        """
+        raise NotImplementedError
+
+
+class MaximumLikelihoodRanker(SequenceRanker):
+    """
+    Select the sample with the highest log probabilities, penalized using either
+    a simple length normalization or Google NMT paper's length penalty
+    """
+
+    def __init__(self, length_penalty: Optional[float]):
+        self.length_penalty = length_penalty
+
+    def rank(self, tokens: List[List[Tensor]], sum_logprobs: List[List[float]]):
+        def scores(logprobs, lengths):
+            result = []
+            for logprob, length in zip(logprobs, lengths):
+                if self.length_penalty is None:
+                    penalty = length
+                else:
+                    # from the Google NMT paper
+                    penalty = ((5 + length) / 6) ** self.length_penalty
+                result.append(logprob / penalty)
+            return result
+
+        # get the sequence with the highest score
+        lengths = [[len(t) for t in s] for s in tokens]
+        return [np.argmax(scores(p, l)) for p, l in zip(sum_logprobs, lengths)]
+
+
+class TokenDecoder:
+    def reset(self):
+        """Initialize any stateful variables for decoding a new sequence"""
+
+    def update(
+        self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor
+    ) -> Tuple[Tensor, bool]:
+        """Specify how to select the next token, based on the current trace and logits
+
+        Parameters
+        ----------
+        tokens : Tensor, shape = (n_batch, current_sequence_length)
+            all tokens in the context so far, including the prefix and sot_sequence tokens
+
+        logits : Tensor, shape = (n_batch, vocab_size)
+            per-token logits of the probability distribution at the current step
+
+        sum_logprobs : Tensor, shape = (n_batch)
+            cumulative log probabilities for each sequence
+
+        Returns
+        -------
+        tokens : Tensor, shape = (n_batch, current_sequence_length + 1)
+            the tokens, appended with the selected next token
+
+        completed : bool
+            True if all sequences has reached the end of text
+
+        """
+        raise NotImplementedError
+
+    def finalize(
+        self, tokens: Tensor, sum_logprobs: Tensor
+    ) -> Tuple[Sequence[Sequence[Tensor]], List[List[float]]]:
+        """Finalize search and return the final candidate sequences
+
+        Parameters
+        ----------
+        tokens : Tensor, shape = (n_audio, n_group, current_sequence_length)
+            all tokens in the context so far, including the prefix and sot_sequence
+
+        sum_logprobs : Tensor, shape = (n_audio, n_group)
+            cumulative log probabilities for each sequence
+
+        Returns
+        -------
+        tokens : Sequence[Sequence[Tensor]], length = n_audio
+            sequence of Tensors containing candidate token sequences, for each audio input
+
+        sum_logprobs : List[List[float]], length = n_audio
+            sequence of cumulative log probabilities corresponding to the above
+
+        """
+        raise NotImplementedError
+
+
+class GreedyDecoder(TokenDecoder):
+    def __init__(self, temperature: float, eot: int):
+        self.temperature = temperature
+        self.eot = eot
+
+    def update(
+        self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor
+    ) -> Tuple[Tensor, bool]:
+        if self.temperature == 0:
+            next_tokens = logits.argmax(dim=-1)
+        else:
+            next_tokens = Categorical(logits=logits / self.temperature).sample()
+
+        logprobs = F.log_softmax(logits.float(), dim=-1)
+        current_logprobs = logprobs[torch.arange(logprobs.shape[0]), next_tokens]
+        sum_logprobs += current_logprobs * (tokens[:, -1] != self.eot)
+
+        next_tokens[tokens[:, -1] == self.eot] = self.eot
+        tokens = torch.cat([tokens, next_tokens[:, None]], dim=-1)
+
+        completed = (tokens[:, -1] == self.eot).all()
+        return tokens, completed
+
+    def finalize(self, tokens: Tensor, sum_logprobs: Tensor):
+        # make sure each sequence has at least one EOT token at the end
+        tokens = F.pad(tokens, (0, 1), value=self.eot)
+        return tokens, sum_logprobs.tolist()
+
+
+class BeamSearchDecoder(TokenDecoder):
+    def __init__(
+        self,
+        beam_size: int,
+        eot: int,
+        inference: Inference,
+        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 = None
+
+        assert (
+            self.max_candidates > 0
+        ), f"Invalid beam size ({beam_size}) or patience ({patience})"
+
+    def reset(self):
+        self.finished_sequences = None
+
+    def update(
+        self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor
+    ) -> Tuple[Tensor, bool]:
+        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:  # for the first update
+            self.finished_sequences = [{} for _ in range(n_audio)]
+
+        logprobs = F.log_softmax(logits.float(), dim=-1)
+        next_tokens, source_indices, finished_sequences = [], [], []
+        for i in range(n_audio):
+            scores, sources, finished = {}, {}, {}
+
+            # STEP 1: calculate the cumulative log probabilities for possible candidates
+            for j in range(self.beam_size):
+                idx = i * self.beam_size + j
+                prefix = tokens[idx].tolist()
+                for logprob, token in zip(*logprobs[idx].topk(self.beam_size + 1)):
+                    new_logprob = (sum_logprobs[idx] + logprob).item()
+                    sequence = tuple(prefix + [token.item()])
+                    scores[sequence] = new_logprob
+                    sources[sequence] = idx
+
+            # STEP 2: rank the candidates and keep the top beam_size sequences for each audio
+            saved = 0
+            for sequence in sorted(scores, key=scores.get, reverse=True):
+                if sequence[-1] == self.eot:
+                    finished[sequence] = scores[sequence]
+                else:
+                    sum_logprobs[len(next_tokens)] = scores[sequence]
+                    next_tokens.append(sequence)
+                    source_indices.append(sources[sequence])
+
+                    saved += 1
+                    if saved == self.beam_size:
+                        break
+
+            finished_sequences.append(finished)
+
+        tokens = torch.tensor(next_tokens, device=tokens.device)
+        self.inference.rearrange_kv_cache(source_indices)
+
+        # add newly finished sequences to self.finished_sequences
+        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  # the candidate list is full
+                previously_finished[seq] = newly_finished[seq]
+
+        # mark as completed if all audio has enough number of samples
+        completed = all(
+            len(sequences) >= self.max_candidates
+            for sequences in self.finished_sequences
+        )
+        return tokens, completed
+
+    def finalize(self, preceding_tokens: Tensor, sum_logprobs: Tensor):
+        # collect all finished sequences, including patience, and add unfinished ones if not enough
+        sum_logprobs = sum_logprobs.cpu()
+        for i, sequences in enumerate(self.finished_sequences):
+            if (
+                len(sequences) < self.beam_size
+            ):  # when not enough sequences are finished
+                for j in list(np.argsort(sum_logprobs[i]))[::-1]:
+                    sequence = preceding_tokens[i, j].tolist() + [self.eot]
+                    sequences[tuple(sequence)] = sum_logprobs[i][j].item()
+                    if len(sequences) >= self.beam_size:
+                        break
+
+        tokens: List[List[Tensor]] = [
+            [torch.tensor(seq) for seq in sequences.keys()]
+            for sequences in self.finished_sequences
+        ]
+        sum_logprobs: List[List[float]] = [
+            list(sequences.values()) for sequences in self.finished_sequences
+        ]
+        return tokens, sum_logprobs
+
+
+class LogitFilter:
+    def apply(self, logits: Tensor, tokens: Tensor) -> None:
+        """Apply any filtering or masking to logits in-place
+
+        Parameters
+        ----------
+        logits : Tensor, shape = (n_batch, vocab_size)
+            per-token logits of the probability distribution at the current step
+
+        tokens : Tensor, shape = (n_batch, current_sequence_length)
+            all tokens in the context so far, including the prefix and sot_sequence tokens
+
+        """
+        raise NotImplementedError
+
+
+class SuppressBlank(LogitFilter):
+    def __init__(self, tokenizer: Tokenizer, sample_begin: int):
+        self.tokenizer = tokenizer
+        self.sample_begin = sample_begin
+
+    def apply(self, logits: Tensor, tokens: Tensor):
+        if tokens.shape[1] == self.sample_begin:
+            logits[:, self.tokenizer.encode(" ") + [self.tokenizer.eot]] = -np.inf
+
+
+class SuppressTokens(LogitFilter):
+    def __init__(self, suppress_tokens: Sequence[int]):
+        self.suppress_tokens = list(suppress_tokens)
+
+    def apply(self, logits: Tensor, tokens: Tensor):
+        logits[:, self.suppress_tokens] = -np.inf
+
+
+class ApplyTimestampRules(LogitFilter):
+    def __init__(
+        self,
+        tokenizer: Tokenizer,
+        sample_begin: int,
+        max_initial_timestamp_index: Optional[int],
+    ):
+        self.tokenizer = tokenizer
+        self.sample_begin = sample_begin
+        self.max_initial_timestamp_index = max_initial_timestamp_index
+
+    def apply(self, logits: Tensor, tokens: Tensor):
+        # suppress <|notimestamps|> which is handled by without_timestamps
+        if self.tokenizer.no_timestamps is not None:
+            logits[:, self.tokenizer.no_timestamps] = -np.inf
+
+        # timestamps have to appear in pairs, except directly before EOT; mask logits accordingly
+        for k in range(tokens.shape[0]):
+            sampled_tokens = tokens[k, self.sample_begin :]
+            seq = [t for t in sampled_tokens.tolist()]
+            last_was_timestamp = (
+                len(seq) >= 1 and seq[-1] >= self.tokenizer.timestamp_begin
+            )
+            penultimate_was_timestamp = (
+                len(seq) < 2 or seq[-2] >= self.tokenizer.timestamp_begin
+            )
+
+            if last_was_timestamp:
+                if penultimate_was_timestamp:  # has to be non-timestamp
+                    logits[k, self.tokenizer.timestamp_begin :] = -np.inf
+                else:  # cannot be normal text tokens
+                    logits[k, : self.tokenizer.eot] = -np.inf
+
+            timestamps = sampled_tokens[
+                sampled_tokens.ge(self.tokenizer.timestamp_begin)
+            ]
+            if timestamps.numel() > 0:
+                # timestamps shouldn't decrease; forbid timestamp tokens smaller than the last
+                # also force each segment to have a nonzero length, to prevent infinite looping
+                if last_was_timestamp and not penultimate_was_timestamp:
+                    timestamp_last = timestamps[-1]
+                else:
+                    timestamp_last = timestamps[-1] + 1
+                logits[k, self.tokenizer.timestamp_begin : timestamp_last] = -np.inf
+
+        if tokens.shape[1] == self.sample_begin:
+            # suppress generating non-timestamp tokens at the beginning
+            logits[:, : self.tokenizer.timestamp_begin] = -np.inf
+
+            # apply the `max_initial_timestamp` option
+            if self.max_initial_timestamp_index is not None:
+                last_allowed = (
+                    self.tokenizer.timestamp_begin + self.max_initial_timestamp_index
+                )
+                logits[:, last_allowed + 1 :] = -np.inf
+
+        # if sum of probability over timestamps is above any other token, sample timestamp
+        logprobs = F.log_softmax(logits.float(), dim=-1)
+        for k in range(tokens.shape[0]):
+            timestamp_logprob = logprobs[k, self.tokenizer.timestamp_begin :].logsumexp(
+                dim=-1
+            )
+            max_text_token_logprob = logprobs[k, : self.tokenizer.timestamp_begin].max()
+            if timestamp_logprob > max_text_token_logprob:
+                logits[k, : self.tokenizer.timestamp_begin] = -np.inf
+
+
+class DecodingTask:
+    inference: Inference
+    sequence_ranker: SequenceRanker
+    decoder: TokenDecoder
+    logit_filters: List[LogitFilter]
+
+    def __init__(self, model: "Whisper", options: DecodingOptions):
+        self.model = model
+
+        language = options.language or "en"
+        tokenizer = get_tokenizer(
+            model.is_multilingual,
+            num_languages=model.num_languages,
+            language=language,
+            task=options.task,
+        )
+        self.tokenizer: Tokenizer = tokenizer
+        self.options: DecodingOptions = self._verify_options(options)
+
+        self.n_group: int = options.beam_size or options.best_of or 1
+        self.n_ctx: int = model.dims.n_text_ctx
+        self.sample_len: int = options.sample_len or model.dims.n_text_ctx // 2
+
+        self.sot_sequence: Tuple[int] = tokenizer.sot_sequence
+        if self.options.without_timestamps:
+            self.sot_sequence = tokenizer.sot_sequence_including_notimestamps
+
+        self.initial_tokens: Tuple[int] = self._get_initial_tokens()
+        self.sample_begin: int = len(self.initial_tokens)
+        self.sot_index: int = self.initial_tokens.index(tokenizer.sot)
+
+        # inference: implements the forward pass through the decoder, including kv caching
+        self.inference = PyTorchInference(model, len(self.initial_tokens))
+
+        # sequence ranker: implements how to rank a group of sampled sequences
+        self.sequence_ranker = MaximumLikelihoodRanker(options.length_penalty)
+
+        # decoder: implements how to select the next tokens, given the autoregressive distribution
+        if options.beam_size is not None:
+            self.decoder = BeamSearchDecoder(
+                options.beam_size, tokenizer.eot, self.inference, options.patience
+            )
+        else:
+            self.decoder = GreedyDecoder(options.temperature, tokenizer.eot)
+
+        # logit filters: applies various rules to suppress or penalize certain tokens
+        self.logit_filters = []
+        if self.options.suppress_blank:
+            self.logit_filters.append(SuppressBlank(self.tokenizer, self.sample_begin))
+        if self.options.suppress_tokens:
+            self.logit_filters.append(SuppressTokens(self._get_suppress_tokens()))
+        if not options.without_timestamps:
+            precision = CHUNK_LENGTH / model.dims.n_audio_ctx  # usually 0.02 seconds
+            max_initial_timestamp_index = None
+            if options.max_initial_timestamp:
+                max_initial_timestamp_index = round(
+                    self.options.max_initial_timestamp / precision
+                )
+            self.logit_filters.append(
+                ApplyTimestampRules(
+                    tokenizer, self.sample_begin, max_initial_timestamp_index
+                )
+            )
+
+    def _verify_options(self, options: DecodingOptions) -> DecodingOptions:
+        if options.beam_size is not None and options.best_of is not None:
+            raise ValueError("beam_size and best_of can't be given together")
+        if options.temperature == 0:
+            if options.best_of is not None:
+                raise ValueError("best_of with greedy sampling (T=0) is not compatible")
+        if options.patience is not None and options.beam_size is None:
+            raise ValueError("patience requires beam_size to be given")
+        if options.length_penalty is not None and not (
+            0 <= options.length_penalty <= 1
+        ):
+            raise ValueError("length_penalty (alpha) should be a value between 0 and 1")
+
+        return options
+
+    def _get_initial_tokens(self) -> Tuple[int]:
+        tokens = list(self.sot_sequence)
+
+        if prefix := self.options.prefix:
+            prefix_tokens = (
+                self.tokenizer.encode(" " + prefix.strip())
+                if isinstance(prefix, str)
+                else prefix
+            )
+            if self.sample_len is not None:
+                max_prefix_len = self.n_ctx // 2 - self.sample_len
+                prefix_tokens = prefix_tokens[-max_prefix_len:]
+            tokens = tokens + prefix_tokens
+
+        if prompt := self.options.prompt:
+            prompt_tokens = (
+                self.tokenizer.encode(" " + prompt.strip())
+                if isinstance(prompt, str)
+                else prompt
+            )
+            tokens = (
+                [self.tokenizer.sot_prev]
+                + prompt_tokens[-(self.n_ctx // 2 - 1) :]
+                + tokens
+            )
+
+        return tuple(tokens)
+
+    def _get_suppress_tokens(self) -> Tuple[int]:
+        suppress_tokens = self.options.suppress_tokens
+
+        if isinstance(suppress_tokens, str):
+            suppress_tokens = [int(t) for t in suppress_tokens.split(",")]
+
+        if -1 in suppress_tokens:
+            suppress_tokens = [t for t in suppress_tokens if t >= 0]
+            suppress_tokens.extend(self.tokenizer.non_speech_tokens)
+        elif suppress_tokens is None or len(suppress_tokens) == 0:
+            suppress_tokens = []  # interpret empty string as an empty list
+        else:
+            assert isinstance(suppress_tokens, list), "suppress_tokens must be a list"
+
+        suppress_tokens.extend(
+            [
+                self.tokenizer.transcribe,
+                self.tokenizer.translate,
+                self.tokenizer.sot,
+                self.tokenizer.sot_prev,
+                self.tokenizer.sot_lm,
+            ]
+        )
+        if self.tokenizer.no_speech is not None:
+            # no-speech probability is collected separately
+            suppress_tokens.append(self.tokenizer.no_speech)
+
+        return tuple(sorted(set(suppress_tokens)))
+
+    def _get_audio_features(self, mel: Tensor):
+        if self.options.fp16:
+            mel = mel.half()
+
+        if mel.shape[-2:] == (
+            self.model.dims.n_audio_ctx,
+            self.model.dims.n_audio_state,
+        ):
+            # encoded audio features are given; skip audio encoding
+            audio_features = mel
+        else:
+            audio_features = self.model.encoder(mel)
+
+        if audio_features.dtype != (
+            torch.float16 if self.options.fp16 else torch.float32
+        ):
+            return TypeError(
+                f"audio_features has an incorrect dtype: {audio_features.dtype}"
+            )
+
+        return audio_features
+
+    def _detect_language(self, audio_features: Tensor, tokens: Tensor):
+        languages = [self.options.language] * audio_features.shape[0]
+        lang_probs = None
+
+        if self.options.language is None or self.options.task == "lang_id":
+            lang_tokens, lang_probs = self.model.detect_language(
+                audio_features, self.tokenizer
+            )
+            languages = [max(probs, key=probs.get) for probs in lang_probs]
+            if self.options.language is None:
+                tokens[:, self.sot_index + 1] = lang_tokens  # write language tokens
+
+        return languages, lang_probs
+
+    def _main_loop(self, audio_features: Tensor, tokens: Tensor):
+        n_batch = tokens.shape[0]
+        sum_logprobs: Tensor = torch.zeros(n_batch, device=audio_features.device)
+        no_speech_probs = [np.nan] * n_batch
+
+        try:
+            for i in range(self.sample_len):
+                logits = self.inference.logits(tokens, audio_features)
+
+                if (
+                    i == 0 and self.tokenizer.no_speech is not None
+                ):  # save no_speech_probs
+                    probs_at_sot = logits[:, self.sot_index].float().softmax(dim=-1)
+                    no_speech_probs = probs_at_sot[:, self.tokenizer.no_speech].tolist()
+
+                # now we need to consider the logits at the last token only
+                logits = logits[:, -1]
+
+                # apply the logit filters, e.g. for suppressing or applying penalty to
+                for logit_filter in self.logit_filters:
+                    logit_filter.apply(logits, tokens)
+
+                # expand the tokens tensor with the selected next tokens
+                tokens, completed = self.decoder.update(tokens, logits, sum_logprobs)
+
+                if completed or tokens.shape[-1] > self.n_ctx:
+                    break
+        finally:
+            self.inference.cleanup_caching()
+
+        return tokens, sum_logprobs, no_speech_probs
+
+    @torch.no_grad()
+    def run(self, mel: Tensor) -> List[DecodingResult]:
+        self.decoder.reset()
+        tokenizer: Tokenizer = self.tokenizer
+        n_audio: int = mel.shape[0]
+
+        audio_features: Tensor = self._get_audio_features(mel)  # encoder forward pass
+        tokens: Tensor = torch.tensor([self.initial_tokens]).repeat(n_audio, 1)
+
+        # detect language if requested, overwriting the language token
+        languages, language_probs = self._detect_language(audio_features, tokens)
+        if self.options.task == "lang_id":
+            return [
+                DecodingResult(
+                    audio_features=features, language=language, language_probs=probs
+                )
+                for features, language, probs in zip(
+                    audio_features, languages, language_probs
+                )
+            ]
+
+        # repeat text tensors by the group size, for beam search or best-of-n sampling
+        tokens = tokens.repeat_interleave(self.n_group, dim=0).to(audio_features.device)
+
+        # call the main sampling loop
+        tokens, sum_logprobs, no_speech_probs = self._main_loop(audio_features, tokens)
+
+        # reshape the tensors to have (n_audio, n_group) as the first two dimensions
+        audio_features = audio_features[:: self.n_group]
+        no_speech_probs = no_speech_probs[:: self.n_group]
+        assert audio_features.shape[0] == len(no_speech_probs) == n_audio
+
+        tokens = tokens.reshape(n_audio, self.n_group, -1)
+        sum_logprobs = sum_logprobs.reshape(n_audio, self.n_group)
+
+        # get the final candidates for each group, and slice between the first sampled token and EOT
+        tokens, sum_logprobs = self.decoder.finalize(tokens, sum_logprobs)
+        tokens: List[List[Tensor]] = [
+            [t[self.sample_begin : (t == tokenizer.eot).nonzero()[0, 0]] for t in s]
+            for s in tokens
+        ]
+
+        # select the top-ranked sample in each group
+        selected = self.sequence_ranker.rank(tokens, sum_logprobs)
+        tokens: List[List[int]] = [t[i].tolist() for i, t in zip(selected, tokens)]
+        texts: List[str] = [tokenizer.decode(t).strip() for t in tokens]
+
+        sum_logprobs: List[float] = [lp[i] for i, lp in zip(selected, sum_logprobs)]
+        avg_logprobs: List[float] = [
+            lp / (len(t) + 1) for t, lp in zip(tokens, sum_logprobs)
+        ]
+
+        fields = (
+            texts,
+            languages,
+            tokens,
+            audio_features,
+            avg_logprobs,
+            no_speech_probs,
+        )
+        if len(set(map(len, fields))) != 1:
+            raise RuntimeError(f"inconsistent result lengths: {list(map(len, fields))}")
+
+        return [
+            DecodingResult(
+                audio_features=features,
+                language=language,
+                tokens=tokens,
+                text=text,
+                avg_logprob=avg_logprob,
+                no_speech_prob=no_speech_prob,
+                temperature=self.options.temperature,
+                compression_ratio=compression_ratio(text),
+            )
+            for text, language, tokens, features, avg_logprob, no_speech_prob in zip(
+                *fields
+            )
+        ]
+
+
+@torch.no_grad()
+def decode(
+    model: "Whisper",
+    mel: Tensor,
+    options: DecodingOptions = DecodingOptions(),
+    **kwargs,
+) -> Union[DecodingResult, List[DecodingResult]]:
+    """
+    Performs decoding of 30-second audio segment(s), provided as Mel spectrogram(s).
+
+    Parameters
+    ----------
+    model: Whisper
+        the Whisper model instance
+
+    mel: torch.Tensor, shape = (80, 3000) or (*, 80, 3000)
+        A tensor containing the Mel spectrogram(s)
+
+    options: DecodingOptions
+        A dataclass that contains all necessary options for decoding 30-second segments
+
+    Returns
+    -------
+    result: Union[DecodingResult, List[DecodingResult]]
+        The result(s) of decoding contained in `DecodingResult` dataclass instance(s)
+    """
+    if single := mel.ndim == 2:
+        mel = mel.unsqueeze(0)
+
+    if kwargs:
+        options = replace(options, **kwargs)
+
+    result = DecodingTask(model, options).run(mel)
+
+    return result[0] if single else result

whisperlivekit/simul_whisper/whisper/model.py

@@ -0,0 +1,348 @@
+import base64
+import gzip
+from contextlib import contextmanager
+from dataclasses import dataclass
+from typing import Dict, Iterable, Optional, Tuple
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import Tensor, nn
+
+from .decoding import decode as decode_function
+from .decoding import detect_language as detect_language_function
+from .transcribe import transcribe as transcribe_function
+
+try:
+    from torch.nn.functional import scaled_dot_product_attention
+
+    SDPA_AVAILABLE = True
+except (ImportError, RuntimeError, OSError):
+    scaled_dot_product_attention = None
+    SDPA_AVAILABLE = False
+
+
+@dataclass
+class ModelDimensions:
+    n_mels: int
+    n_audio_ctx: int
+    n_audio_state: int
+    n_audio_head: int
+    n_audio_layer: int
+    n_vocab: int
+    n_text_ctx: int
+    n_text_state: int
+    n_text_head: int
+    n_text_layer: int
+
+
+class LayerNorm(nn.LayerNorm):
+    def forward(self, x: Tensor) -> Tensor:
+        return super().forward(x.float()).type(x.dtype)
+
+
+class Linear(nn.Linear):
+    def forward(self, x: Tensor) -> Tensor:
+        return F.linear(
+            x,
+            self.weight.to(x.dtype),
+            None if self.bias is None else self.bias.to(x.dtype),
+        )
+
+
+class Conv1d(nn.Conv1d):
+    def _conv_forward(
+        self, x: Tensor, weight: Tensor, bias: Optional[Tensor]
+    ) -> Tensor:
+        return super()._conv_forward(
+            x, weight.to(x.dtype), None if bias is None else bias.to(x.dtype)
+        )
+
+
+def sinusoids(length, channels, max_timescale=10000):
+    """Returns sinusoids for positional embedding"""
+    assert channels % 2 == 0
+    log_timescale_increment = np.log(max_timescale) / (channels // 2 - 1)
+    inv_timescales = torch.exp(-log_timescale_increment * torch.arange(channels // 2))
+    scaled_time = torch.arange(length)[:, np.newaxis] * inv_timescales[np.newaxis, :]
+    return torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=1)
+
+
+@contextmanager
+def disable_sdpa():
+    prev_state = MultiHeadAttention.use_sdpa
+    try:
+        MultiHeadAttention.use_sdpa = False
+        yield
+    finally:
+        MultiHeadAttention.use_sdpa = prev_state
+
+
+class MultiHeadAttention(nn.Module):
+    use_sdpa = False  # Disable SDPA to ensure qk is always computed for hooks
+
+    def __init__(self, n_state: int, n_head: int, cache_id: str = ""):
+        super().__init__()
+        self.n_head = n_head
+        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"
+
+    def forward(
+        self,
+        x: Tensor,
+        xa: Optional[Tensor] = None,
+        mask: Optional[Tensor] = None,
+        kv_cache: Optional[dict] = None,
+    ):
+        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)
+        else:
+            # for cross-attention, calculate keys and values once and reuse in subsequent calls.
+            k = kv_cache[self.key]
+            v = kv_cache[self.value]
+
+        wv, qk = self.qkv_attention(q, k, v, mask)
+        return self.out(wv), qk
+
+    def qkv_attention(
+        self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor] = None
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        n_batch, n_ctx, n_state = q.shape
+        scale = (n_state // self.n_head) ** -0.25
+        q = q.view(*q.shape[:2], self.n_head, -1).permute(0, 2, 1, 3)
+        k = k.view(*k.shape[:2], self.n_head, -1).permute(0, 2, 1, 3)
+        v = v.view(*v.shape[:2], self.n_head, -1).permute(0, 2, 1, 3)
+
+        if SDPA_AVAILABLE and MultiHeadAttention.use_sdpa:
+            a = scaled_dot_product_attention(
+                q, k, v, is_causal=mask is not None and n_ctx > 1
+            )
+            out = a.permute(0, 2, 1, 3).flatten(start_dim=2)
+            qk = None
+        else:
+            qk = (q * scale) @ (k * scale).transpose(-1, -2)
+            if mask is not None:
+                qk = qk + mask[:n_ctx, :n_ctx]
+            qk = qk.float()
+
+            w = F.softmax(qk, dim=-1).to(q.dtype)
+            out = (w @ v).permute(0, 2, 1, 3).flatten(start_dim=2)
+            qk = qk.detach()
+
+        return out, qk
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(self, n_state: int, n_head: int, cross_attention: bool = False, cache_id: str = ""):
+        super().__init__()
+
+        self.attn = MultiHeadAttention(n_state, n_head, cache_id=f"{cache_id}_self_attn")
+        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
+        )
+        self.cross_attn_ln = LayerNorm(n_state) if cross_attention else None
+
+        n_mlp = n_state * 4
+        self.mlp = nn.Sequential(
+            Linear(n_state, n_mlp), nn.GELU(), Linear(n_mlp, n_state)
+        )
+        self.mlp_ln = LayerNorm(n_state)
+
+    def forward(
+        self,
+        x: Tensor,
+        xa: Optional[Tensor] = None,
+        mask: Optional[Tensor] = None,
+        kv_cache: Optional[dict] = None,
+    ):
+        x = x + self.attn(self.attn_ln(x), mask=mask, kv_cache=kv_cache)[0]
+        if self.cross_attn:
+            x = x + self.cross_attn(self.cross_attn_ln(x), xa, kv_cache=kv_cache)[0]
+        x = x + self.mlp(self.mlp_ln(x))
+        return x
+
+
+class AudioEncoder(nn.Module):
+    def __init__(
+        self, n_mels: int, n_ctx: int, n_state: int, n_head: int, n_layer: int
+    ):
+        super().__init__()
+        self.conv1 = Conv1d(n_mels, n_state, kernel_size=3, padding=1)
+        self.conv2 = Conv1d(n_state, n_state, kernel_size=3, stride=2, padding=1)
+        self.register_buffer("positional_embedding", sinusoids(n_ctx, n_state))
+
+        self.blocks: Iterable[ResidualAttentionBlock] = nn.ModuleList(
+            [ResidualAttentionBlock(n_state, n_head, cache_id=f"enc_layer{i}") for i in range(n_layer)]
+        )
+        self.ln_post = LayerNorm(n_state)
+
+    def forward(self, x: Tensor):
+        """
+        x : torch.Tensor, shape = (batch_size, n_mels, n_ctx)
+            the mel spectrogram of the audio
+        """
+        x = F.gelu(self.conv1(x))
+        x = F.gelu(self.conv2(x))
+        x = x.permute(0, 2, 1)
+
+        assert x.shape[1:] == self.positional_embedding.shape, "incorrect audio shape"
+        x = (x + self.positional_embedding).to(x.dtype)
+
+        for block in self.blocks:
+            x = block(x)
+
+        x = self.ln_post(x)
+        return x
+
+
+class TextDecoder(nn.Module):
+    def __init__(
+        self, n_vocab: int, n_ctx: int, n_state: int, n_head: int, n_layer: int
+    ):
+        super().__init__()
+
+        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}")
+                for i in range(n_layer)
+            ]
+        )
+        self.ln = LayerNorm(n_state)
+
+        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):
+        """
+        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
+        """
+        offset = next(iter(kv_cache.values())).shape[1] if kv_cache else 0
+        x = (
+            self.token_embedding(x)
+            + self.positional_embedding[offset : offset + x.shape[-1]]
+        )
+        x = x.to(xa.dtype)
+
+        for block in self.blocks:
+            x = block(x, xa, mask=self.mask, kv_cache=kv_cache)
+
+        x = self.ln(x)
+        logits = (
+            x @ torch.transpose(self.token_embedding.weight.to(x.dtype), 0, 1)
+        ).float()
+
+        return logits
+
+
+class Whisper(nn.Module):
+    def __init__(self, dims: ModelDimensions):
+        super().__init__()
+        self.dims = dims
+        self.encoder = AudioEncoder(
+            self.dims.n_mels,
+            self.dims.n_audio_ctx,
+            self.dims.n_audio_state,
+            self.dims.n_audio_head,
+            self.dims.n_audio_layer,
+        )
+        self.decoder = TextDecoder(
+            self.dims.n_vocab,
+            self.dims.n_text_ctx,
+            self.dims.n_text_state,
+            self.dims.n_text_head,
+            self.dims.n_text_layer,
+        )
+        # use the last half among the decoder layers for time alignment by default;
+        # to use a specific set of heads, see `set_alignment_heads()` below.
+        all_heads = torch.zeros(
+            self.dims.n_text_layer, self.dims.n_text_head, dtype=torch.bool
+        )
+        all_heads[self.dims.n_text_layer // 2 :] = True
+        self.register_buffer("alignment_heads", all_heads.to_sparse(), persistent=False)
+
+    def set_alignment_heads(self, dump: bytes):
+        array = np.frombuffer(
+            gzip.decompress(base64.b85decode(dump)), dtype=bool
+        ).copy()
+        mask = torch.from_numpy(array).reshape(
+            self.dims.n_text_layer, self.dims.n_text_head
+        )
+        self.register_buffer("alignment_heads", mask.to_sparse(), persistent=False)
+
+    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 forward(
+        self, mel: torch.Tensor, tokens: torch.Tensor
+    ) -> Dict[str, torch.Tensor]:
+        return self.decoder(tokens, self.encoder(mel))
+
+    @property
+    def device(self):
+        return next(self.parameters()).device
+
+    @property
+    def is_multilingual(self):
+        return self.dims.n_vocab >= 51865
+
+    @property
+    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/simul_whisper/whisper/normalizers/__init__.py

@@ -0,0 +1,2 @@
+from .basic import BasicTextNormalizer as BasicTextNormalizer
+from .english import EnglishTextNormalizer as EnglishTextNormalizer

whisperlivekit/simul_whisper/whisper/normalizers/basic.py

@@ -0,0 +1,80 @@
+import re
+import unicodedata
+
+import regex
+
+# non-ASCII letters that are not separated by "NFKD" normalization
+ADDITIONAL_DIACRITICS = {
+    "œ": "oe",
+    "Œ": "OE",
+    "ø": "o",
+    "Ø": "O",
+    "æ": "ae",
+    "Æ": "AE",
+    "ß": "ss",
+    "ẞ": "SS",
+    "đ": "d",
+    "Đ": "D",
+    "ð": "d",
+    "Ð": "D",
+    "þ": "th",
+    "Þ": "th",
+    "ł": "l",
+    "Ł": "L",
+}
+
+
+def remove_symbols_and_diacritics(s: str, keep=""):
+    """
+    Replace any other markers, symbols, and punctuations with a space,
+    and drop any diacritics (category 'Mn' and some manual mappings)
+    """
+    return "".join(
+        (
+            c
+            if c in keep
+            else (
+                ADDITIONAL_DIACRITICS[c]
+                if c in ADDITIONAL_DIACRITICS
+                else (
+                    ""
+                    if unicodedata.category(c) == "Mn"
+                    else " " if unicodedata.category(c)[0] in "MSP" else c
+                )
+            )
+        )
+        for c in unicodedata.normalize("NFKD", s)
+    )
+
+
+def remove_symbols(s: str):
+    """
+    Replace any other markers, symbols, punctuations with a space, keeping diacritics
+    """
+    return "".join(
+        " " if unicodedata.category(c)[0] in "MSP" else c
+        for c in unicodedata.normalize("NFKC", s)
+    )
+
+
+class BasicTextNormalizer:
+    def __init__(self, remove_diacritics: bool = False, split_letters: bool = False):
+        self.clean = (
+            remove_symbols_and_diacritics if remove_diacritics else remove_symbols
+        )
+        self.split_letters = split_letters
+
+    def __call__(self, s: str):
+        s = s.lower()
+        s = re.sub(r"[<\[][^>\]]*[>\]]", "", s)  # remove words between brackets
+        s = re.sub(r"\(([^)]+?)\)", "", s)  # remove words between parenthesis
+        s = self.clean(s).lower()
+
+        if self.split_letters:
+            s = " ".join(regex.findall(r"\X", s, regex.U))
+
+        s = re.sub(
+            r"\s+", " ", s
+        )  # replace any successive whitespace characters with a space
+
+        return s

whisperlivekit/simul_whisper/whisper/normalizers/english.py

@@ -0,0 +1,550 @@
+import json
+import os
+import re
+from fractions import Fraction
+from typing import Iterator, List, Match, Optional, Union
+
+from more_itertools import windowed
+
+from .basic import remove_symbols_and_diacritics
+
+
+class EnglishNumberNormalizer:
+    """
+    Convert any spelled-out numbers into arabic numbers, while handling:
+
+    - remove any commas
+    - keep the suffixes such as: `1960s`, `274th`, `32nd`, etc.
+    - spell out currency symbols after the number. e.g. `$20 million` -> `20000000 dollars`
+    - spell out `one` and `ones`
+    - interpret successive single-digit numbers as nominal: `one oh one` -> `101`
+    """
+
+    def __init__(self):
+        super().__init__()
+
+        self.zeros = {"o", "oh", "zero"}
+        self.ones = {
+            name: i
+            for i, name in enumerate(
+                [
+                    "one",
+                    "two",
+                    "three",
+                    "four",
+                    "five",
+                    "six",
+                    "seven",
+                    "eight",
+                    "nine",
+                    "ten",
+                    "eleven",
+                    "twelve",
+                    "thirteen",
+                    "fourteen",
+                    "fifteen",
+                    "sixteen",
+                    "seventeen",
+                    "eighteen",
+                    "nineteen",
+                ],
+                start=1,
+            )
+        }
+        self.ones_plural = {
+            "sixes" if name == "six" else name + "s": (value, "s")
+            for name, value in self.ones.items()
+        }
+        self.ones_ordinal = {
+            "zeroth": (0, "th"),
+            "first": (1, "st"),
+            "second": (2, "nd"),
+            "third": (3, "rd"),
+            "fifth": (5, "th"),
+            "twelfth": (12, "th"),
+            **{
+                name + ("h" if name.endswith("t") else "th"): (value, "th")
+                for name, value in self.ones.items()
+                if value > 3 and value != 5 and value != 12
+            },
+        }
+        self.ones_suffixed = {**self.ones_plural, **self.ones_ordinal}
+
+        self.tens = {
+            "twenty": 20,
+            "thirty": 30,
+            "forty": 40,
+            "fifty": 50,
+            "sixty": 60,
+            "seventy": 70,
+            "eighty": 80,
+            "ninety": 90,
+        }
+        self.tens_plural = {
+            name.replace("y", "ies"): (value, "s") for name, value in self.tens.items()
+        }
+        self.tens_ordinal = {
+            name.replace("y", "ieth"): (value, "th")
+            for name, value in self.tens.items()
+        }
+        self.tens_suffixed = {**self.tens_plural, **self.tens_ordinal}
+
+        self.multipliers = {
+            "hundred": 100,
+            "thousand": 1_000,
+            "million": 1_000_000,
+            "billion": 1_000_000_000,
+            "trillion": 1_000_000_000_000,
+            "quadrillion": 1_000_000_000_000_000,
+            "quintillion": 1_000_000_000_000_000_000,
+            "sextillion": 1_000_000_000_000_000_000_000,
+            "septillion": 1_000_000_000_000_000_000_000_000,
+            "octillion": 1_000_000_000_000_000_000_000_000_000,
+            "nonillion": 1_000_000_000_000_000_000_000_000_000_000,
+            "decillion": 1_000_000_000_000_000_000_000_000_000_000_000,
+        }
+        self.multipliers_plural = {
+            name + "s": (value, "s") for name, value in self.multipliers.items()
+        }
+        self.multipliers_ordinal = {
+            name + "th": (value, "th") for name, value in self.multipliers.items()
+        }
+        self.multipliers_suffixed = {
+            **self.multipliers_plural,
+            **self.multipliers_ordinal,
+        }
+        self.decimals = {*self.ones, *self.tens, *self.zeros}
+
+        self.preceding_prefixers = {
+            "minus": "-",
+            "negative": "-",
+            "plus": "+",
+            "positive": "+",
+        }
+        self.following_prefixers = {
+            "pound": "£",
+            "pounds": "£",
+            "euro": "€",
+            "euros": "€",
+            "dollar": "$",
+            "dollars": "$",
+            "cent": "¢",
+            "cents": "¢",
+        }
+        self.prefixes = set(
+            list(self.preceding_prefixers.values())
+            + list(self.following_prefixers.values())
+        )
+        self.suffixers = {
+            "per": {"cent": "%"},
+            "percent": "%",
+        }
+        self.specials = {"and", "double", "triple", "point"}
+
+        self.words = set(
+            [
+                key
+                for mapping in [
+                    self.zeros,
+                    self.ones,
+                    self.ones_suffixed,
+                    self.tens,
+                    self.tens_suffixed,
+                    self.multipliers,
+                    self.multipliers_suffixed,
+                    self.preceding_prefixers,
+                    self.following_prefixers,
+                    self.suffixers,
+                    self.specials,
+                ]
+                for key in mapping
+            ]
+        )
+        self.literal_words = {"one", "ones"}
+
+    def process_words(self, words: List[str]) -> Iterator[str]:
+        prefix: Optional[str] = None
+        value: Optional[Union[str, int]] = None
+        skip = False
+
+        def to_fraction(s: str):
+            try:
+                return Fraction(s)
+            except ValueError:
+                return None
+
+        def output(result: Union[str, int]):
+            nonlocal prefix, value
+            result = str(result)
+            if prefix is not None:
+                result = prefix + result
+            value = None
+            prefix = None
+            return result
+
+        if len(words) == 0:
+            return
+
+        for prev, current, next in windowed([None] + words + [None], 3):
+            if skip:
+                skip = False
+                continue
+
+            next_is_numeric = next is not None and re.match(r"^\d+(\.\d+)?$", next)
+            has_prefix = current[0] in self.prefixes
+            current_without_prefix = current[1:] if has_prefix else current
+            if re.match(r"^\d+(\.\d+)?$", current_without_prefix):
+                # arabic numbers (potentially with signs and fractions)
+                f = to_fraction(current_without_prefix)
+                assert f is not None
+                if value is not None:
+                    if isinstance(value, str) and value.endswith("."):
+                        # concatenate decimals / ip address components
+                        value = str(value) + str(current)
+                        continue
+                    else:
+                        yield output(value)
+
+                prefix = current[0] if has_prefix else prefix
+                if f.denominator == 1:
+                    value = f.numerator  # store integers as int
+                else:
+                    value = current_without_prefix
+            elif current not in self.words:
+                # non-numeric words
+                if value is not None:
+                    yield output(value)
+                yield output(current)
+            elif current in self.zeros:
+                value = str(value or "") + "0"
+            elif current in self.ones:
+                ones = self.ones[current]
+
+                if value is None:
+                    value = ones
+                elif isinstance(value, str) or prev in self.ones:
+                    if (
+                        prev in self.tens and ones < 10
+                    ):  # replace the last zero with the digit
+                        assert value[-1] == "0"
+                        value = value[:-1] + str(ones)
+                    else:
+                        value = str(value) + str(ones)
+                elif ones < 10:
+                    if value % 10 == 0:
+                        value += ones
+                    else:
+                        value = str(value) + str(ones)
+                else:  # eleven to nineteen
+                    if value % 100 == 0:
+                        value += ones
+                    else:
+                        value = str(value) + str(ones)
+            elif current in self.ones_suffixed:
+                # ordinal or cardinal; yield the number right away
+                ones, suffix = self.ones_suffixed[current]
+                if value is None:
+                    yield output(str(ones) + suffix)
+                elif isinstance(value, str) or prev in self.ones:
+                    if prev in self.tens and ones < 10:
+                        assert value[-1] == "0"
+                        yield output(value[:-1] + str(ones) + suffix)
+                    else:
+                        yield output(str(value) + str(ones) + suffix)
+                elif ones < 10:
+                    if value % 10 == 0:
+                        yield output(str(value + ones) + suffix)
+                    else:
+                        yield output(str(value) + str(ones) + suffix)
+                else:  # eleven to nineteen
+                    if value % 100 == 0:
+                        yield output(str(value + ones) + suffix)
+                    else:
+                        yield output(str(value) + str(ones) + suffix)
+                value = None
+            elif current in self.tens:
+                tens = self.tens[current]
+                if value is None:
+                    value = tens
+                elif isinstance(value, str):
+                    value = str(value) + str(tens)
+                else:
+                    if value % 100 == 0:
+                        value += tens
+                    else:
+                        value = str(value) + str(tens)
+            elif current in self.tens_suffixed:
+                # ordinal or cardinal; yield the number right away
+                tens, suffix = self.tens_suffixed[current]
+                if value is None:
+                    yield output(str(tens) + suffix)
+                elif isinstance(value, str):
+                    yield output(str(value) + str(tens) + suffix)
+                else:
+                    if value % 100 == 0:
+                        yield output(str(value + tens) + suffix)
+                    else:
+                        yield output(str(value) + str(tens) + suffix)
+            elif current in self.multipliers:
+                multiplier = self.multipliers[current]
+                if value is None:
+                    value = multiplier
+                elif isinstance(value, str) or value == 0:
+                    f = to_fraction(value)
+                    p = f * multiplier if f is not None else None
+                    if f is not None and p.denominator == 1:
+                        value = p.numerator
+                    else:
+                        yield output(value)
+                        value = multiplier
+                else:
+                    before = value // 1000 * 1000
+                    residual = value % 1000
+                    value = before + residual * multiplier
+            elif current in self.multipliers_suffixed:
+                multiplier, suffix = self.multipliers_suffixed[current]
+                if value is None:
+                    yield output(str(multiplier) + suffix)
+                elif isinstance(value, str):
+                    f = to_fraction(value)
+                    p = f * multiplier if f is not None else None
+                    if f is not None and p.denominator == 1:
+                        yield output(str(p.numerator) + suffix)
+                    else:
+                        yield output(value)
+                        yield output(str(multiplier) + suffix)
+                else:  # int
+                    before = value // 1000 * 1000
+                    residual = value % 1000
+                    value = before + residual * multiplier
+                    yield output(str(value) + suffix)
+                value = None
+            elif current in self.preceding_prefixers:
+                # apply prefix (positive, minus, etc.) if it precedes a number
+                if value is not None:
+                    yield output(value)
+
+                if next in self.words or next_is_numeric:
+                    prefix = self.preceding_prefixers[current]
+                else:
+                    yield output(current)
+            elif current in self.following_prefixers:
+                # apply prefix (dollars, cents, etc.) only after a number
+                if value is not None:
+                    prefix = self.following_prefixers[current]
+                    yield output(value)
+                else:
+                    yield output(current)
+            elif current in self.suffixers:
+                # apply suffix symbols (percent -> '%')
+                if value is not None:
+                    suffix = self.suffixers[current]
+                    if isinstance(suffix, dict):
+                        if next in suffix:
+                            yield output(str(value) + suffix[next])
+                            skip = True
+                        else:
+                            yield output(value)
+                            yield output(current)
+                    else:
+                        yield output(str(value) + suffix)
+                else:
+                    yield output(current)
+            elif current in self.specials:
+                if next not in self.words and not next_is_numeric:
+                    # apply special handling only if the next word can be numeric
+                    if value is not None:
+                        yield output(value)
+                    yield output(current)
+                elif current == "and":
+                    # ignore "and" after hundreds, thousands, etc.
+                    if prev not in self.multipliers:
+                        if value is not None:
+                            yield output(value)
+                        yield output(current)
+                elif current == "double" or current == "triple":
+                    if next in self.ones or next in self.zeros:
+                        repeats = 2 if current == "double" else 3
+                        ones = self.ones.get(next, 0)
+                        value = str(value or "") + str(ones) * repeats
+                        skip = True
+                    else:
+                        if value is not None:
+                            yield output(value)
+                        yield output(current)
+                elif current == "point":
+                    if next in self.decimals or next_is_numeric:
+                        value = str(value or "") + "."
+                else:
+                    # should all have been covered at this point
+                    raise ValueError(f"Unexpected token: {current}")
+            else:
+                # all should have been covered at this point
+                raise ValueError(f"Unexpected token: {current}")
+
+        if value is not None:
+            yield output(value)
+
+    def preprocess(self, s: str):
+        # replace "<number> and a half" with "<number> point five"
+        results = []
+
+        segments = re.split(r"\band\s+a\s+half\b", s)
+        for i, segment in enumerate(segments):
+            if len(segment.strip()) == 0:
+                continue
+            if i == len(segments) - 1:
+                results.append(segment)
+            else:
+                results.append(segment)
+                last_word = segment.rsplit(maxsplit=2)[-1]
+                if last_word in self.decimals or last_word in self.multipliers:
+                    results.append("point five")
+                else:
+                    results.append("and a half")
+
+        s = " ".join(results)
+
+        # put a space at number/letter boundary
+        s = re.sub(r"([a-z])([0-9])", r"\1 \2", s)
+        s = re.sub(r"([0-9])([a-z])", r"\1 \2", s)
+
+        # but remove spaces which could be a suffix
+        s = re.sub(r"([0-9])\s+(st|nd|rd|th|s)\b", r"\1\2", s)
+
+        return s
+
+    def postprocess(self, s: str):
+        def combine_cents(m: Match):
+            try:
+                currency = m.group(1)
+                integer = m.group(2)
+                cents = int(m.group(3))
+                return f"{currency}{integer}.{cents:02d}"
+            except ValueError:
+                return m.string
+
+        def extract_cents(m: Match):
+            try:
+                return f"¢{int(m.group(1))}"
+            except ValueError:
+                return m.string
+
+        # apply currency postprocessing; "$2 and ¢7" -> "$2.07"
+        s = re.sub(r"([€£$])([0-9]+) (?:and )?¢([0-9]{1,2})\b", combine_cents, s)
+        s = re.sub(r"[€£$]0.([0-9]{1,2})\b", extract_cents, s)
+
+        # write "one(s)" instead of "1(s)", just for the readability
+        s = re.sub(r"\b1(s?)\b", r"one\1", s)
+
+        return s
+
+    def __call__(self, s: str):
+        s = self.preprocess(s)
+        s = " ".join(word for word in self.process_words(s.split()) if word is not None)
+        s = self.postprocess(s)
+
+        return s
+
+
+class EnglishSpellingNormalizer:
+    """
+    Applies British-American spelling mappings as listed in [1].
+
+    [1] https://www.tysto.com/uk-us-spelling-list.html
+    """
+
+    def __init__(self):
+        mapping_path = os.path.join(os.path.dirname(__file__), "english.json")
+        self.mapping = json.load(open(mapping_path))
+
+    def __call__(self, s: str):
+        return " ".join(self.mapping.get(word, word) for word in s.split())
+
+
+class EnglishTextNormalizer:
+    def __init__(self):
+        self.ignore_patterns = r"\b(hmm|mm|mhm|mmm|uh|um)\b"
+        self.replacers = {
+            # common contractions
+            r"\bwon't\b": "will not",
+            r"\bcan't\b": "can not",
+            r"\blet's\b": "let us",
+            r"\bain't\b": "aint",
+            r"\by'all\b": "you all",
+            r"\bwanna\b": "want to",
+            r"\bgotta\b": "got to",
+            r"\bgonna\b": "going to",
+            r"\bi'ma\b": "i am going to",
+            r"\bimma\b": "i am going to",
+            r"\bwoulda\b": "would have",
+            r"\bcoulda\b": "could have",
+            r"\bshoulda\b": "should have",
+            r"\bma'am\b": "madam",
+            # contractions in titles/prefixes
+            r"\bmr\b": "mister ",
+            r"\bmrs\b": "missus ",
+            r"\bst\b": "saint ",
+            r"\bdr\b": "doctor ",
+            r"\bprof\b": "professor ",
+            r"\bcapt\b": "captain ",
+            r"\bgov\b": "governor ",
+            r"\bald\b": "alderman ",
+            r"\bgen\b": "general ",
+            r"\bsen\b": "senator ",
+            r"\brep\b": "representative ",
+            r"\bpres\b": "president ",
+            r"\brev\b": "reverend ",
+            r"\bhon\b": "honorable ",
+            r"\basst\b": "assistant ",
+            r"\bassoc\b": "associate ",
+            r"\blt\b": "lieutenant ",
+            r"\bcol\b": "colonel ",
+            r"\bjr\b": "junior ",
+            r"\bsr\b": "senior ",
+            r"\besq\b": "esquire ",
+            # prefect tenses, ideally it should be any past participles, but it's harder..
+            r"'d been\b": " had been",
+            r"'s been\b": " has been",
+            r"'d gone\b": " had gone",
+            r"'s gone\b": " has gone",
+            r"'d done\b": " had done",  # "'s done" is ambiguous
+            r"'s got\b": " has got",
+            # general contractions
+            r"n't\b": " not",
+            r"'re\b": " are",
+            r"'s\b": " is",
+            r"'d\b": " would",
+            r"'ll\b": " will",
+            r"'t\b": " not",
+            r"'ve\b": " have",
+            r"'m\b": " am",
+        }
+        self.standardize_numbers = EnglishNumberNormalizer()
+        self.standardize_spellings = EnglishSpellingNormalizer()
+
+    def __call__(self, s: str):
+        s = s.lower()
+
+        s = re.sub(r"[<\[][^>\]]*[>\]]", "", s)  # remove words between brackets
+        s = re.sub(r"\(([^)]+?)\)", "", s)  # remove words between parenthesis
+        s = re.sub(self.ignore_patterns, "", s)
+        s = re.sub(r"\s+'", "'", s)  # when there's a space before an apostrophe
+
+        for pattern, replacement in self.replacers.items():
+            s = re.sub(pattern, replacement, s)
+
+        s = re.sub(r"(\d),(\d)", r"\1\2", s)  # remove commas between digits
+        s = re.sub(r"\.([^0-9]|$)", r" \1", s)  # remove periods not followed by numbers
+        s = remove_symbols_and_diacritics(s, keep=".%$¢€£")  # keep numeric symbols
+
+        s = self.standardize_numbers(s)
+        s = self.standardize_spellings(s)
+
+        # now remove prefix/suffix symbols that are not preceded/followed by numbers
+        s = re.sub(r"[.$¢€£]([^0-9])", r" \1", s)
+        s = re.sub(r"([^0-9])%", r"\1 ", s)
+
+        s = re.sub(r"\s+", " ", s)  # replace any successive whitespaces with a space
+
+        return s

whisperlivekit/simul_whisper/whisper/timing.py

@@ -0,0 +1,388 @@
+import itertools
+import subprocess
+import warnings
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, List
+
+import numba
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+from .audio import HOP_LENGTH, SAMPLE_RATE, TOKENS_PER_SECOND
+from .tokenizer import Tokenizer
+
+if TYPE_CHECKING:
+    from .model import Whisper
+
+
+def median_filter(x: torch.Tensor, filter_width: int):
+    """Apply a median filter of width `filter_width` along the last dimension of `x`"""
+    pad_width = filter_width // 2
+    if x.shape[-1] <= pad_width:
+        # F.pad requires the padding width to be smaller than the input dimension
+        return x
+
+    if (ndim := x.ndim) <= 2:
+        # `F.pad` does not support 1D or 2D inputs for reflect padding but supports 3D and 4D
+        x = x[None, None, :]
+
+    assert (
+        filter_width > 0 and filter_width % 2 == 1
+    ), "`filter_width` should be an odd number"
+
+    result = None
+    x = F.pad(x, (filter_width // 2, filter_width // 2, 0, 0), mode="reflect")
+    if x.is_cuda:
+        try:
+            from .triton_ops import median_filter_cuda
+
+            result = median_filter_cuda(x, filter_width)
+        except (RuntimeError, subprocess.CalledProcessError):
+            warnings.warn(
+                "Failed to launch Triton kernels, likely due to missing CUDA toolkit; "
+                "falling back to a slower median kernel implementation..."
+            )
+
+    if result is None:
+        # sort() is faster than torch.median (https://github.com/pytorch/pytorch/issues/51450)
+        result = x.unfold(-1, filter_width, 1).sort()[0][..., filter_width // 2]
+
+    if ndim <= 2:
+        result = result[0, 0]
+
+    return result
+
+
+@numba.jit(nopython=True)
+def backtrace(trace: np.ndarray):
+    i = trace.shape[0] - 1
+    j = trace.shape[1] - 1
+    trace[0, :] = 2
+    trace[:, 0] = 1
+
+    result = []
+    while i > 0 or j > 0:
+        result.append((i - 1, j - 1))
+
+        if trace[i, j] == 0:
+            i -= 1
+            j -= 1
+        elif trace[i, j] == 1:
+            i -= 1
+        elif trace[i, j] == 2:
+            j -= 1
+        else:
+            raise ValueError("Unexpected trace[i, j]")
+
+    result = np.array(result)
+    return result[::-1, :].T
+
+
+@numba.jit(nopython=True, parallel=True)
+def dtw_cpu(x: np.ndarray):
+    N, M = x.shape
+    cost = np.ones((N + 1, M + 1), dtype=np.float32) * np.inf
+    trace = -np.ones((N + 1, M + 1), dtype=np.float32)
+
+    cost[0, 0] = 0
+    for j in range(1, M + 1):
+        for i in range(1, N + 1):
+            c0 = cost[i - 1, j - 1]
+            c1 = cost[i - 1, j]
+            c2 = cost[i, j - 1]
+
+            if c0 < c1 and c0 < c2:
+                c, t = c0, 0
+            elif c1 < c0 and c1 < c2:
+                c, t = c1, 1
+            else:
+                c, t = c2, 2
+
+            cost[i, j] = x[i - 1, j - 1] + c
+            trace[i, j] = t
+
+    return backtrace(trace)
+
+
+def dtw_cuda(x, BLOCK_SIZE=1024):
+    from .triton_ops import dtw_kernel
+
+    M, N = x.shape
+    assert M < BLOCK_SIZE, f"M should be smaller than {BLOCK_SIZE=}"
+
+    x_skew = (
+        F.pad(x, (0, M + 1), value=np.inf).flatten()[: M * (N + M)].reshape(M, N + M)
+    )
+    x_skew = x_skew.T.contiguous()
+    cost = torch.ones(N + M + 2, M + 2) * np.inf
+    cost[0, 0] = 0
+    cost = cost.to(x.device)
+    trace = torch.zeros_like(cost, dtype=torch.int32)
+
+    dtw_kernel[(1,)](
+        cost,
+        trace,
+        x_skew,
+        x_skew.stride(0),
+        cost.stride(0),
+        trace.stride(0),
+        N,
+        M,
+        BLOCK_SIZE=BLOCK_SIZE,
+    )
+
+    trace = trace.T.flatten()[: (M + 1) * (M + N + 3)].reshape(M + 1, M + N + 3)[
+        :, : N + 1
+    ]
+    return backtrace(trace.cpu().numpy())
+
+
+def dtw(x: torch.Tensor) -> np.ndarray:
+    if x.is_cuda:
+        try:
+            return dtw_cuda(x)
+        except (RuntimeError, subprocess.CalledProcessError):
+            warnings.warn(
+                "Failed to launch Triton kernels, likely due to missing CUDA toolkit; "
+                "falling back to a slower DTW implementation..."
+            )
+
+    return dtw_cpu(x.double().cpu().numpy())
+
+
+@dataclass
+class WordTiming:
+    word: str
+    tokens: List[int]
+    start: float
+    end: float
+    probability: float
+
+
+def find_alignment(
+    model: "Whisper",
+    tokenizer: Tokenizer,
+    text_tokens: List[int],
+    mel: torch.Tensor,
+    num_frames: int,
+    *,
+    medfilt_width: int = 7,
+    qk_scale: float = 1.0,
+) -> List[WordTiming]:
+    if len(text_tokens) == 0:
+        return []
+
+    tokens = torch.tensor(
+        [
+            *tokenizer.sot_sequence,
+            tokenizer.no_timestamps,
+            *text_tokens,
+            tokenizer.eot,
+        ]
+    ).to(model.device)
+
+    # install hooks on the cross attention layers to retrieve the attention weights
+    QKs = [None] * model.dims.n_text_layer
+    hooks = [
+        block.cross_attn.register_forward_hook(
+            lambda _, ins, outs, index=i: QKs.__setitem__(index, outs[-1][0])
+        )
+        for i, block in enumerate(model.decoder.blocks)
+    ]
+
+    from .model import disable_sdpa
+
+    with torch.no_grad(), disable_sdpa():
+        logits = model(mel.unsqueeze(0), tokens.unsqueeze(0))[0]
+        sampled_logits = logits[len(tokenizer.sot_sequence) :, : tokenizer.eot]
+        token_probs = sampled_logits.softmax(dim=-1)
+        text_token_probs = token_probs[np.arange(len(text_tokens)), text_tokens]
+        text_token_probs = text_token_probs.tolist()
+
+    for hook in hooks:
+        hook.remove()
+
+    # heads * tokens * frames
+    weights = torch.stack([QKs[_l][_h] for _l, _h in model.alignment_heads.indices().T])
+    weights = weights[:, :, : num_frames // 2]
+    weights = (weights * qk_scale).softmax(dim=-1)
+    std, mean = torch.std_mean(weights, dim=-2, keepdim=True, unbiased=False)
+    weights = (weights - mean) / std
+    weights = median_filter(weights, medfilt_width)
+
+    matrix = weights.mean(axis=0)
+    matrix = matrix[len(tokenizer.sot_sequence) : -1]
+    text_indices, time_indices = dtw(-matrix)
+
+    words, word_tokens = tokenizer.split_to_word_tokens(text_tokens + [tokenizer.eot])
+    if len(word_tokens) <= 1:
+        # return on eot only
+        # >>> np.pad([], (1, 0))
+        # array([0.])
+        # This results in crashes when we lookup jump_times with float, like
+        # IndexError: arrays used as indices must be of integer (or boolean) type
+        return []
+    word_boundaries = np.pad(np.cumsum([len(t) for t in word_tokens[:-1]]), (1, 0))
+
+    jumps = np.pad(np.diff(text_indices), (1, 0), constant_values=1).astype(bool)
+    jump_times = time_indices[jumps] / TOKENS_PER_SECOND
+    start_times = jump_times[word_boundaries[:-1]]
+    end_times = jump_times[word_boundaries[1:]]
+    word_probabilities = [
+        np.mean(text_token_probs[i:j])
+        for i, j in zip(word_boundaries[:-1], word_boundaries[1:])
+    ]
+
+    return [
+        WordTiming(word, tokens, start, end, probability)
+        for word, tokens, start, end, probability in zip(
+            words, word_tokens, start_times, end_times, word_probabilities
+        )
+    ]
+
+
+def merge_punctuations(alignment: List[WordTiming], prepended: str, appended: str):
+    # merge prepended punctuations
+    i = len(alignment) - 2
+    j = len(alignment) - 1
+    while i >= 0:
+        previous = alignment[i]
+        following = alignment[j]
+        if previous.word.startswith(" ") and previous.word.strip() in prepended:
+            # prepend it to the following word
+            following.word = previous.word + following.word
+            following.tokens = previous.tokens + following.tokens
+            previous.word = ""
+            previous.tokens = []
+        else:
+            j = i
+        i -= 1
+
+    # merge appended punctuations
+    i = 0
+    j = 1
+    while j < len(alignment):
+        previous = alignment[i]
+        following = alignment[j]
+        if not previous.word.endswith(" ") and following.word in appended:
+            # append it to the previous word
+            previous.word = previous.word + following.word
+            previous.tokens = previous.tokens + following.tokens
+            following.word = ""
+            following.tokens = []
+        else:
+            i = j
+        j += 1
+
+
+def add_word_timestamps(
+    *,
+    segments: List[dict],
+    model: "Whisper",
+    tokenizer: Tokenizer,
+    mel: torch.Tensor,
+    num_frames: int,
+    prepend_punctuations: str = "\"'“¿([{-",
+    append_punctuations: str = "\"'.。,，!！?？:：”)]}、",
+    last_speech_timestamp: float,
+    **kwargs,
+):
+    if len(segments) == 0:
+        return
+
+    text_tokens_per_segment = [
+        [token for token in segment["tokens"] if token < tokenizer.eot]
+        for segment in segments
+    ]
+
+    text_tokens = list(itertools.chain.from_iterable(text_tokens_per_segment))
+    alignment = find_alignment(model, tokenizer, text_tokens, mel, num_frames, **kwargs)
+    word_durations = np.array([t.end - t.start for t in alignment])
+    word_durations = word_durations[word_durations.nonzero()]
+    median_duration = np.median(word_durations) if len(word_durations) > 0 else 0.0
+    median_duration = min(0.7, float(median_duration))
+    max_duration = median_duration * 2
+
+    # hack: truncate long words at sentence boundaries.
+    # a better segmentation algorithm based on VAD should be able to replace this.
+    if len(word_durations) > 0:
+        sentence_end_marks = ".。!！?？"
+        # ensure words at sentence boundaries are not longer than twice the median word duration.
+        for i in range(1, len(alignment)):
+            if alignment[i].end - alignment[i].start > max_duration:
+                if alignment[i].word in sentence_end_marks:
+                    alignment[i].end = alignment[i].start + max_duration
+                elif alignment[i - 1].word in sentence_end_marks:
+                    alignment[i].start = alignment[i].end - max_duration
+
+    merge_punctuations(alignment, prepend_punctuations, append_punctuations)
+
+    time_offset = segments[0]["seek"] * HOP_LENGTH / SAMPLE_RATE
+    word_index = 0
+
+    for segment, text_tokens in zip(segments, text_tokens_per_segment):
+        saved_tokens = 0
+        words = []
+
+        while word_index < len(alignment) and saved_tokens < len(text_tokens):
+            timing = alignment[word_index]
+
+            if timing.word:
+                words.append(
+                    dict(
+                        word=timing.word,
+                        start=round(time_offset + timing.start, 2),
+                        end=round(time_offset + timing.end, 2),
+                        probability=timing.probability,
+                    )
+                )
+
+            saved_tokens += len(timing.tokens)
+            word_index += 1
+
+        # hack: truncate long words at segment boundaries.
+        # a better segmentation algorithm based on VAD should be able to replace this.
+        if len(words) > 0:
+            # ensure the first and second word after a pause is not longer than
+            # twice the median word duration.
+            if words[0]["end"] - last_speech_timestamp > median_duration * 4 and (
+                words[0]["end"] - words[0]["start"] > max_duration
+                or (
+                    len(words) > 1
+                    and words[1]["end"] - words[0]["start"] > max_duration * 2
+                )
+            ):
+                if (
+                    len(words) > 1
+                    and words[1]["end"] - words[1]["start"] > max_duration
+                ):
+                    boundary = max(words[1]["end"] / 2, words[1]["end"] - max_duration)
+                    words[0]["end"] = words[1]["start"] = boundary
+                words[0]["start"] = max(0, words[0]["end"] - max_duration)
+
+            # prefer the segment-level start timestamp if the first word is too long.
+            if (
+                segment["start"] < words[0]["end"]
+                and segment["start"] - 0.5 > words[0]["start"]
+            ):
+                words[0]["start"] = max(
+                    0, min(words[0]["end"] - median_duration, segment["start"])
+                )
+            else:
+                segment["start"] = words[0]["start"]
+
+            # prefer the segment-level end timestamp if the last word is too long.
+            if (
+                segment["end"] > words[-1]["start"]
+                and segment["end"] + 0.5 < words[-1]["end"]
+            ):
+                words[-1]["end"] = max(
+                    words[-1]["start"] + median_duration, segment["end"]
+                )
+            else:
+                segment["end"] = words[-1]["end"]
+
+            last_speech_timestamp = segment["end"]
+
+        segment["words"] = words

whisperlivekit/simul_whisper/whisper/tokenizer.py

@@ -0,0 +1,395 @@
+import base64
+import os
+import string
+from dataclasses import dataclass, field
+from functools import cached_property, lru_cache
+from typing import Dict, List, Optional, Tuple
+
+import tiktoken
+
+LANGUAGES = {
+    "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",
+}
+
+# language code lookup by name, with a few language aliases
+TO_LANGUAGE_CODE = {
+    **{language: code for code, language in LANGUAGES.items()},
+    "burmese": "my",
+    "valencian": "ca",
+    "flemish": "nl",
+    "haitian": "ht",
+    "letzeburgesch": "lb",
+    "pushto": "ps",
+    "panjabi": "pa",
+    "moldavian": "ro",
+    "moldovan": "ro",
+    "sinhalese": "si",
+    "castilian": "es",
+    "mandarin": "zh",
+}
+
+
+@dataclass
+class Tokenizer:
+    """A thin wrapper around `tiktoken` providing quick access to special tokens"""
+
+    encoding: tiktoken.Encoding
+    num_languages: int
+    language: Optional[str] = None
+    task: Optional[str] = None
+    sot_sequence: Tuple[int] = ()
+    special_tokens: Dict[str, int] = field(default_factory=dict)
+
+    def __post_init__(self):
+        for special in self.encoding.special_tokens_set:
+            special_token = self.encoding.encode_single_token(special)
+            self.special_tokens[special] = special_token
+
+        sot: int = self.special_tokens["<|startoftranscript|>"]
+        translate: int = self.special_tokens["<|translate|>"]
+        transcribe: int = self.special_tokens["<|transcribe|>"]
+
+        langs = tuple(LANGUAGES.keys())[: self.num_languages]
+        sot_sequence = [sot]
+        if self.language is not None:
+            sot_sequence.append(sot + 1 + langs.index(self.language))
+        if self.task is not None:
+            task_token: int = transcribe if self.task == "transcribe" else translate
+            sot_sequence.append(task_token)
+
+        self.sot_sequence = tuple(sot_sequence)
+
+    def encode(self, text, **kwargs):
+        return self.encoding.encode(text, **kwargs)
+
+    def decode(self, token_ids: List[int], **kwargs) -> str:
+        token_ids = [t for t in token_ids if t < self.timestamp_begin]
+        return self.encoding.decode(token_ids, **kwargs)
+
+    def decode_with_timestamps(self, token_ids: List[int], **kwargs) -> str:
+        """
+        Timestamp tokens are above other special tokens' id range and are ignored by `decode()`.
+        This method decodes given tokens with timestamps tokens annotated, e.g. "<|1.08|>".
+        """
+        return self.encoding.decode(token_ids, **kwargs)
+
+    @cached_property
+    def eot(self) -> int:
+        return self.encoding.eot_token
+
+    @cached_property
+    def transcribe(self) -> int:
+        return self.special_tokens["<|transcribe|>"]
+
+    @cached_property
+    def translate(self) -> int:
+        return self.special_tokens["<|translate|>"]
+
+    @cached_property
+    def sot(self) -> int:
+        return self.special_tokens["<|startoftranscript|>"]
+
+    @cached_property
+    def sot_lm(self) -> int:
+        return self.special_tokens["<|startoflm|>"]
+
+    @cached_property
+    def sot_prev(self) -> int:
+        return self.special_tokens["<|startofprev|>"]
+
+    @cached_property
+    def no_speech(self) -> int:
+        return self.special_tokens["<|nospeech|>"]
+
+    @cached_property
+    def no_timestamps(self) -> int:
+        return self.special_tokens["<|notimestamps|>"]
+
+    @cached_property
+    def timestamp_begin(self) -> int:
+        return self.special_tokens["<|0.00|>"]
+
+    @cached_property
+    def language_token(self) -> int:
+        """Returns the token id corresponding to the value of the `language` field"""
+        if self.language is None:
+            raise ValueError("This tokenizer does not have language token configured")
+
+        return self.to_language_token(self.language)
+
+    def to_language_token(self, language):
+        if token := self.special_tokens.get(f"<|{language}|>", None):
+            return token
+
+        raise KeyError(f"Language {language} not found in tokenizer.")
+
+    @cached_property
+    def all_language_tokens(self) -> Tuple[int]:
+        result = []
+        for token, token_id in self.special_tokens.items():
+            if token.strip("<|>") in LANGUAGES:
+                result.append(token_id)
+        return tuple(result)[: self.num_languages]
+
+    @cached_property
+    def all_language_codes(self) -> Tuple[str]:
+        return tuple(self.decode([_l]).strip("<|>") for _l in self.all_language_tokens)
+
+    @cached_property
+    def sot_sequence_including_notimestamps(self) -> Tuple[int]:
+        return tuple(list(self.sot_sequence) + [self.no_timestamps])
+
+    @cached_property
+    def non_speech_tokens(self) -> Tuple[int]:
+        """
+        Returns the list of tokens to suppress in order to avoid any speaker tags or non-speech
+        annotations, to prevent sampling texts that are not actually spoken in the audio, e.g.
+
+        - ♪♪♪
+        - ( SPEAKING FOREIGN LANGUAGE )
+        - [DAVID] Hey there,
+
+        keeping basic punctuations like commas, periods, question marks, exclamation points, etc.
+        """
+        symbols = list('"#()*+/:;<=>@[\\]^_`{|}~「」『』')
+        symbols += (
+            "<< >> <<< >>> -- --- -( -[ (' (\" (( )) ((( ))) [[ ]] {{ }} ♪♪ ♪♪♪".split()
+        )
+
+        # symbols that may be a single token or multiple tokens depending on the tokenizer.
+        # In case they're multiple tokens, suppress the first token, which is safe because:
+        # These are between U+2640 and U+267F miscellaneous symbols that are okay to suppress
+        # in generations, and in the 3-byte UTF-8 representation they share the first two bytes.
+        miscellaneous = set("♩♪♫♬♭♮♯")
+        assert all(0x2640 <= ord(c) <= 0x267F for c in miscellaneous)
+
+        # allow hyphens "-" and single quotes "'" between words, but not at the beginning of a word
+        result = {self.encoding.encode(" -")[0], self.encoding.encode(" '")[0]}
+        for symbol in symbols + list(miscellaneous):
+            for tokens in [
+                self.encoding.encode(symbol),
+                self.encoding.encode(" " + symbol),
+            ]:
+                if len(tokens) == 1 or symbol in miscellaneous:
+                    result.add(tokens[0])
+
+        return tuple(sorted(result))
+
+    def split_to_word_tokens(self, tokens: List[int]):
+        if self.language in {"zh", "ja", "th", "lo", "my", "yue"}:
+            # These languages don't typically use spaces, so it is difficult to split words
+            # without morpheme analysis. Here, we instead split words at any
+            # position where the tokens are decoded as valid unicode points
+            return self.split_tokens_on_unicode(tokens)
+
+        return self.split_tokens_on_spaces(tokens)
+
+    def split_tokens_on_unicode(self, tokens: List[int]):
+        decoded_full = self.decode_with_timestamps(tokens)
+        replacement_char = "\ufffd"
+
+        words = []
+        word_tokens = []
+        current_tokens = []
+        unicode_offset = 0
+
+        for token in tokens:
+            current_tokens.append(token)
+            decoded = self.decode_with_timestamps(current_tokens)
+
+            if (
+                replacement_char not in decoded
+                or decoded_full[unicode_offset + decoded.index(replacement_char)]
+                == replacement_char
+            ):
+                words.append(decoded)
+                word_tokens.append(current_tokens)
+                current_tokens = []
+                unicode_offset += len(decoded)
+
+        return words, word_tokens
+
+    def split_tokens_on_spaces(self, tokens: List[int]):
+        subwords, subword_tokens_list = self.split_tokens_on_unicode(tokens)
+        words = []
+        word_tokens = []
+
+        for subword, subword_tokens in zip(subwords, subword_tokens_list):
+            special = subword_tokens[0] >= self.eot
+            with_space = subword.startswith(" ")
+            punctuation = subword.strip() in string.punctuation
+            if special or with_space or punctuation or len(words) == 0:
+                words.append(subword)
+                word_tokens.append(subword_tokens)
+            else:
+                words[-1] = words[-1] + subword
+                word_tokens[-1].extend(subword_tokens)
+
+        return words, word_tokens
+
+
+@lru_cache(maxsize=None)
+def get_encoding(name: str = "gpt2", num_languages: int = 99):
+    vocab_path = os.path.join(os.path.dirname(__file__), "assets", f"{name}.tiktoken")
+    ranks = {
+        base64.b64decode(token): int(rank)
+        for token, rank in (line.split() for line in open(vocab_path) if line)
+    }
+    n_vocab = len(ranks)
+    special_tokens = {}
+
+    specials = [
+        "<|endoftext|>",
+        "<|startoftranscript|>",
+        *[f"<|{lang}|>" for lang in list(LANGUAGES.keys())[:num_languages]],
+        "<|translate|>",
+        "<|transcribe|>",
+        "<|startoflm|>",
+        "<|startofprev|>",
+        "<|nospeech|>",
+        "<|notimestamps|>",
+        *[f"<|{i * 0.02:.2f}|>" for i in range(1501)],
+    ]
+
+    for token in specials:
+        special_tokens[token] = n_vocab
+        n_vocab += 1
+
+    return tiktoken.Encoding(
+        name=os.path.basename(vocab_path),
+        explicit_n_vocab=n_vocab,
+        pat_str=r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""",
+        mergeable_ranks=ranks,
+        special_tokens=special_tokens,
+    )
+
+
+@lru_cache(maxsize=None)
+def get_tokenizer(
+    multilingual: bool,
+    *,
+    num_languages: int = 99,
+    language: Optional[str] = None,
+    task: Optional[str] = None,  # Literal["transcribe", "translate", None]
+) -> Tokenizer:
+    if language is not None:
+        language = language.lower()
+        if language not in LANGUAGES:
+            if language in TO_LANGUAGE_CODE:
+                language = TO_LANGUAGE_CODE[language]
+            else:
+                raise ValueError(f"Unsupported language: {language}")
+
+    if multilingual:
+        encoding_name = "multilingual"
+        language = language or "en"
+        task = task or "transcribe"
+    else:
+        encoding_name = "gpt2"
+        language = None
+        task = None
+
+    encoding = get_encoding(name=encoding_name, num_languages=num_languages)
+
+    return Tokenizer(
+        encoding=encoding, num_languages=num_languages, language=language, task=task
+    )

whisperlivekit/simul_whisper/whisper/transcribe.py

@@ -0,0 +1,623 @@
+import argparse
+import os
+import traceback
+import warnings
+from typing import TYPE_CHECKING, List, Optional, Tuple, Union
+
+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 .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,
+)
+
+if TYPE_CHECKING:
+    from .model import Whisper
+
+
+def transcribe(
+    model: "Whisper",
+    audio: Union[str, np.ndarray, torch.Tensor],
+    *,
+    verbose: Optional[bool] = None,
+    temperature: Union[float, Tuple[float, ...]] = (0.0, 0.2, 0.4, 0.6, 0.8, 1.0),
+    compression_ratio_threshold: Optional[float] = 2.4,
+    logprob_threshold: Optional[float] = -1.0,
+    no_speech_threshold: Optional[float] = 0.6,
+    condition_on_previous_text: bool = True,
+    initial_prompt: Optional[str] = None,
+    carry_initial_prompt: bool = False,
+    word_timestamps: bool = False,
+    prepend_punctuations: str = "\"'“¿([{-",
+    append_punctuations: str = "\"'.。,，!！?？:：”)]}、",
+    clip_timestamps: Union[str, List[float]] = "0",
+    hallucination_silence_threshold: Optional[float] = None,
+    **decode_options,
+):
+    """
+    Transcribe an audio file using Whisper
+
+    Parameters
+    ----------
+    model: Whisper
+        The Whisper model instance
+
+    audio: Union[str, np.ndarray, torch.Tensor]
+        The path to the audio file to open, or the audio waveform
+
+    verbose: bool
+        Whether to display the text being decoded to the console. If True, displays all the details,
+        If False, displays minimal details. If None, does not display anything
+
+    temperature: Union[float, Tuple[float, ...]]
+        Temperature for sampling. It can be a tuple of temperatures, which will be successively used
+        upon failures according to either `compression_ratio_threshold` or `logprob_threshold`.
+
+    compression_ratio_threshold: float
+        If the gzip compression ratio is above this value, treat as failed
+
+    logprob_threshold: float
+        If the average log probability over sampled tokens is below this value, treat as failed
+
+    no_speech_threshold: float
+        If the no_speech probability is higher than this value AND the average log probability
+        over sampled tokens is below `logprob_threshold`, consider the segment as silent
+
+    condition_on_previous_text: bool
+        if True, the previous output of the model is provided as a prompt for the next window;
+        disabling may make the text inconsistent across windows, but the model becomes less prone to
+        getting stuck in a failure loop, such as repetition looping or timestamps going out of sync.
+
+    word_timestamps: bool
+        Extract word-level timestamps using the cross-attention pattern and dynamic time warping,
+        and include the timestamps for each word in each segment.
+
+    prepend_punctuations: str
+        If word_timestamps is True, merge these punctuation symbols with the next word
+
+    append_punctuations: str
+        If word_timestamps is True, merge these punctuation symbols with the previous word
+
+    initial_prompt: Optional[str]
+        Optional text to provide as a prompt for the first window. This can be used to provide, or
+        "prompt-engineer" a context for transcription, e.g. custom vocabularies or proper nouns
+        to make it more likely to predict those word correctly.
+
+    carry_initial_prompt: bool
+        If carry_initial_prompt is True, `initial_prompt` is prepended to the prompt of each internal
+        `decode()` call. If there is not enough context space at the start of the prompt, it is
+        left-sliced to make space.
+
+    decode_options: dict
+        Keyword arguments to construct `DecodingOptions` instances
+
+    clip_timestamps: Union[str, List[float]]
+        Comma-separated list start,end,start,end,... timestamps (in seconds) of clips to process.
+        The last end timestamp defaults to the end of the file.
+
+    hallucination_silence_threshold: Optional[float]
+        When word_timestamps is True, skip silent periods longer than this threshold (in seconds)
+        when a possible hallucination is detected
+
+    Returns
+    -------
+    A dictionary containing the resulting text ("text") and segment-level details ("segments"), and
+    the spoken language ("language"), which is detected when `decode_options["language"]` is None.
+    """
+    dtype = torch.float16 if decode_options.get("fp16", True) else torch.float32
+    if model.device == torch.device("cpu"):
+        if torch.cuda.is_available():
+            warnings.warn("Performing inference on CPU when CUDA is available")
+        if dtype == torch.float16:
+            warnings.warn("FP16 is not supported on CPU; using FP32 instead")
+            dtype = torch.float32
+
+    if dtype == torch.float32:
+        decode_options["fp16"] = False
+
+    # Pad 30-seconds of silence to the input audio, for slicing
+    mel = log_mel_spectrogram(audio, model.dims.n_mels, padding=N_SAMPLES)
+    content_frames = mel.shape[-1] - N_FRAMES
+    content_duration = float(content_frames * HOP_LENGTH / SAMPLE_RATE)
+
+    if decode_options.get("language", None) is None:
+        if not model.is_multilingual:
+            decode_options["language"] = "en"
+        else:
+            if verbose:
+                print(
+                    "Detecting language using up to the first 30 seconds. Use `--language` to specify the language"
+                )
+            mel_segment = pad_or_trim(mel, N_FRAMES).to(model.device).to(dtype)
+            _, probs = model.detect_language(mel_segment)
+            decode_options["language"] = max(probs, key=probs.get)
+            if verbose is not None:
+                print(
+                    f"Detected language: {LANGUAGES[decode_options['language']].title()}"
+                )
+
+    language: str = decode_options["language"]
+    task: str = decode_options.get("task", "transcribe")
+    tokenizer = get_tokenizer(
+        model.is_multilingual,
+        num_languages=model.num_languages,
+        language=language,
+        task=task,
+    )
+
+    if isinstance(clip_timestamps, str):
+        clip_timestamps = [
+            float(ts) for ts in (clip_timestamps.split(",") if clip_timestamps else [])
+        ]
+    seek_points: List[int] = [round(ts * FRAMES_PER_SECOND) for ts in clip_timestamps]
+    if len(seek_points) == 0:
+        seek_points.append(0)
+    if len(seek_points) % 2 == 1:
+        seek_points.append(content_frames)
+    seek_clips: List[Tuple[int, int]] = list(zip(seek_points[::2], seek_points[1::2]))
+
+    punctuation = "\"'“¿([{-\"'.。,，!！?？:：”)]}、"
+
+    if word_timestamps and task == "translate":
+        warnings.warn("Word-level timestamps on translations may not be reliable.")
+
+    def decode_with_fallback(segment: torch.Tensor) -> DecodingResult:
+        temperatures = (
+            [temperature] if isinstance(temperature, (int, float)) else temperature
+        )
+        decode_result = None
+
+        for t in temperatures:
+            kwargs = {**decode_options}
+            if t > 0:
+                # disable beam_size and patience when t > 0
+                kwargs.pop("beam_size", None)
+                kwargs.pop("patience", None)
+            else:
+                # disable best_of when t == 0
+                kwargs.pop("best_of", None)
+
+            options = DecodingOptions(**kwargs, temperature=t)
+            decode_result = model.decode(segment, options)
+
+            needs_fallback = False
+            if (
+                compression_ratio_threshold is not None
+                and decode_result.compression_ratio > compression_ratio_threshold
+            ):
+                needs_fallback = True  # too repetitive
+            if (
+                logprob_threshold is not None
+                and decode_result.avg_logprob < logprob_threshold
+            ):
+                needs_fallback = True  # average log probability is too low
+            if (
+                no_speech_threshold is not None
+                and decode_result.no_speech_prob > no_speech_threshold
+                and logprob_threshold is not None
+                and decode_result.avg_logprob < logprob_threshold
+            ):
+                needs_fallback = False  # silence
+            if not needs_fallback:
+                break
+
+        return decode_result
+
+    clip_idx = 0
+    seek = seek_clips[clip_idx][0]
+    input_stride = exact_div(
+        N_FRAMES, model.dims.n_audio_ctx
+    )  # mel frames per output token: 2
+    time_precision = (
+        input_stride * HOP_LENGTH / SAMPLE_RATE
+    )  # time per output token: 0.02 (seconds)
+    all_tokens = []
+    all_segments = []
+    prompt_reset_since = 0
+
+    remaining_prompt_length = model.dims.n_text_ctx // 2 - 1
+    if initial_prompt is not None:
+        initial_prompt_tokens = tokenizer.encode(" " + initial_prompt.strip())
+        all_tokens.extend(initial_prompt_tokens)
+        remaining_prompt_length -= len(initial_prompt_tokens)
+    else:
+        initial_prompt_tokens = []
+
+    def new_segment(
+        *, start: float, end: float, tokens: torch.Tensor, result: DecodingResult
+    ):
+        tokens = tokens.tolist()
+        text_tokens = [token for token in tokens if token < tokenizer.eot]
+        return {
+            "seek": seek,
+            "start": start,
+            "end": end,
+            "text": tokenizer.decode(text_tokens),
+            "tokens": tokens,
+            "temperature": result.temperature,
+            "avg_logprob": result.avg_logprob,
+            "compression_ratio": result.compression_ratio,
+            "no_speech_prob": result.no_speech_prob,
+        }
+
+    # show the progress bar when verbose is False (if True, transcribed text will be printed)
+    with tqdm.tqdm(
+        total=content_frames, unit="frames", disable=verbose is not False
+    ) as pbar:
+        last_speech_timestamp = 0.0
+        # NOTE: This loop is obscurely flattened to make the diff readable.
+        # A later commit should turn this into a simpler nested loop.
+        # for seek_clip_start, seek_clip_end in seek_clips:
+        #     while seek < seek_clip_end
+        while clip_idx < len(seek_clips):
+            seek_clip_start, seek_clip_end = seek_clips[clip_idx]
+            if seek < seek_clip_start:
+                seek = seek_clip_start
+            if seek >= seek_clip_end:
+                clip_idx += 1
+                if clip_idx < len(seek_clips):
+                    seek = seek_clips[clip_idx][0]
+                continue
+            time_offset = float(seek * HOP_LENGTH / SAMPLE_RATE)
+            window_end_time = float((seek + N_FRAMES) * HOP_LENGTH / SAMPLE_RATE)
+            segment_size = min(N_FRAMES, content_frames - seek, seek_clip_end - seek)
+            mel_segment = mel[:, seek : seek + segment_size]
+            segment_duration = segment_size * HOP_LENGTH / SAMPLE_RATE
+            mel_segment = pad_or_trim(mel_segment, N_FRAMES).to(model.device).to(dtype)
+
+            if carry_initial_prompt:
+                nignored = max(len(initial_prompt_tokens), prompt_reset_since)
+                remaining_prompt = all_tokens[nignored:][-remaining_prompt_length:]
+                decode_options["prompt"] = initial_prompt_tokens + remaining_prompt
+            else:
+                decode_options["prompt"] = all_tokens[prompt_reset_since:]
+
+            result: DecodingResult = decode_with_fallback(mel_segment)
+            tokens = torch.tensor(result.tokens)
+
+            if no_speech_threshold is not None:
+                # no voice activity check
+                should_skip = result.no_speech_prob > no_speech_threshold
+                if (
+                    logprob_threshold is not None
+                    and result.avg_logprob > logprob_threshold
+                ):
+                    # don't skip if the logprob is high enough, despite the no_speech_prob
+                    should_skip = False
+
+                if should_skip:
+                    seek += segment_size  # fast-forward to the next segment boundary
+                    continue
+
+            previous_seek = seek
+            current_segments = []
+
+            # anomalous words are very long/short/improbable
+            def word_anomaly_score(word: dict) -> float:
+                probability = word.get("probability", 0.0)
+                duration = word["end"] - word["start"]
+                score = 0.0
+                if probability < 0.15:
+                    score += 1.0
+                if duration < 0.133:
+                    score += (0.133 - duration) * 15
+                if duration > 2.0:
+                    score += duration - 2.0
+                return score
+
+            def is_segment_anomaly(segment: Optional[dict]) -> bool:
+                if segment is None or not segment["words"]:
+                    return False
+                words = [w for w in segment["words"] if w["word"] not in punctuation]
+                words = words[:8]
+                score = sum(word_anomaly_score(w) for w in words)
+                return score >= 3 or score + 0.01 >= len(words)
+
+            def next_words_segment(segments: List[dict]) -> Optional[dict]:
+                return next((s for s in segments if s["words"]), None)
+
+            timestamp_tokens: torch.Tensor = tokens.ge(tokenizer.timestamp_begin)
+            single_timestamp_ending = timestamp_tokens[-2:].tolist() == [False, True]
+
+            consecutive = torch.where(timestamp_tokens[:-1] & timestamp_tokens[1:])[0]
+            consecutive.add_(1)
+            if len(consecutive) > 0:
+                # if the output contains two consecutive timestamp tokens
+                slices = consecutive.tolist()
+                if single_timestamp_ending:
+                    slices.append(len(tokens))
+
+                last_slice = 0
+                for current_slice in slices:
+                    sliced_tokens = tokens[last_slice:current_slice]
+                    start_timestamp_pos = (
+                        sliced_tokens[0].item() - tokenizer.timestamp_begin
+                    )
+                    end_timestamp_pos = (
+                        sliced_tokens[-1].item() - tokenizer.timestamp_begin
+                    )
+                    current_segments.append(
+                        new_segment(
+                            start=time_offset + start_timestamp_pos * time_precision,
+                            end=time_offset + end_timestamp_pos * time_precision,
+                            tokens=sliced_tokens,
+                            result=result,
+                        )
+                    )
+                    last_slice = current_slice
+
+                if single_timestamp_ending:
+                    # single timestamp at the end means no speech after the last timestamp.
+                    seek += segment_size
+                else:
+                    # otherwise, ignore the unfinished segment and seek to the last timestamp
+                    last_timestamp_pos = (
+                        tokens[last_slice - 1].item() - tokenizer.timestamp_begin
+                    )
+                    seek += last_timestamp_pos * input_stride
+            else:
+                duration = segment_duration
+                timestamps = tokens[timestamp_tokens.nonzero().flatten()]
+                if (
+                    len(timestamps) > 0
+                    and timestamps[-1].item() != tokenizer.timestamp_begin
+                ):
+                    # no consecutive timestamps but it has a timestamp; use the last one.
+                    last_timestamp_pos = (
+                        timestamps[-1].item() - tokenizer.timestamp_begin
+                    )
+                    duration = last_timestamp_pos * time_precision
+
+                current_segments.append(
+                    new_segment(
+                        start=time_offset,
+                        end=time_offset + duration,
+                        tokens=tokens,
+                        result=result,
+                    )
+                )
+                seek += segment_size
+
+            if word_timestamps:
+                add_word_timestamps(
+                    segments=current_segments,
+                    model=model,
+                    tokenizer=tokenizer,
+                    mel=mel_segment,
+                    num_frames=segment_size,
+                    prepend_punctuations=prepend_punctuations,
+                    append_punctuations=append_punctuations,
+                    last_speech_timestamp=last_speech_timestamp,
+                )
+
+                if not single_timestamp_ending:
+                    last_word_end = get_end(current_segments)
+                    if last_word_end is not None and last_word_end > time_offset:
+                        seek = round(last_word_end * FRAMES_PER_SECOND)
+
+                # skip silence before possible hallucinations
+                if hallucination_silence_threshold is not None:
+                    threshold = hallucination_silence_threshold
+                    if not single_timestamp_ending:
+                        last_word_end = get_end(current_segments)
+                        if last_word_end is not None and last_word_end > time_offset:
+                            remaining_duration = window_end_time - last_word_end
+                            if remaining_duration > threshold:
+                                seek = round(last_word_end * FRAMES_PER_SECOND)
+                            else:
+                                seek = previous_seek + segment_size
+
+                    # if first segment might be a hallucination, skip leading silence
+                    first_segment = next_words_segment(current_segments)
+                    if first_segment is not None and is_segment_anomaly(first_segment):
+                        gap = first_segment["start"] - time_offset
+                        if gap > threshold:
+                            seek = previous_seek + round(gap * FRAMES_PER_SECOND)
+                            continue
+
+                    # skip silence before any possible hallucination that is surrounded
+                    # by silence or more hallucinations
+                    hal_last_end = last_speech_timestamp
+                    for si in range(len(current_segments)):
+                        segment = current_segments[si]
+                        if not segment["words"]:
+                            continue
+                        if is_segment_anomaly(segment):
+                            next_segment = next_words_segment(
+                                current_segments[si + 1 :]
+                            )
+                            if next_segment is not None:
+                                hal_next_start = next_segment["words"][0]["start"]
+                            else:
+                                hal_next_start = time_offset + segment_duration
+                            silence_before = (
+                                segment["start"] - hal_last_end > threshold
+                                or segment["start"] < threshold
+                                or segment["start"] - time_offset < 2.0
+                            )
+                            silence_after = (
+                                hal_next_start - segment["end"] > threshold
+                                or is_segment_anomaly(next_segment)
+                                or window_end_time - segment["end"] < 2.0
+                            )
+                            if silence_before and silence_after:
+                                seek = round(
+                                    max(time_offset + 1, segment["start"])
+                                    * FRAMES_PER_SECOND
+                                )
+                                if content_duration - segment["end"] < threshold:
+                                    seek = content_frames
+                                current_segments[si:] = []
+                                break
+                        hal_last_end = segment["end"]
+
+                last_word_end = get_end(current_segments)
+                if last_word_end is not None:
+                    last_speech_timestamp = last_word_end
+
+            if verbose:
+                for segment in current_segments:
+                    start, end, text = segment["start"], segment["end"], segment["text"]
+                    line = f"[{format_timestamp(start)} --> {format_timestamp(end)}] {text}"
+                    print(make_safe(line))
+
+            # if a segment is instantaneous or does not contain text, clear it
+            for i, segment in enumerate(current_segments):
+                if segment["start"] == segment["end"] or segment["text"].strip() == "":
+                    segment["text"] = ""
+                    segment["tokens"] = []
+                    segment["words"] = []
+
+            all_segments.extend(
+                [
+                    {"id": i, **segment}
+                    for i, segment in enumerate(
+                        current_segments, start=len(all_segments)
+                    )
+                ]
+            )
+            all_tokens.extend(
+                [token for segment in current_segments for token in segment["tokens"]]
+            )
+
+            if not condition_on_previous_text or result.temperature > 0.5:
+                # do not feed the prompt tokens if a high temperature was used
+                prompt_reset_since = len(all_tokens)
+
+            # update progress bar
+            pbar.update(min(content_frames, seek) - previous_seek)
+
+    return dict(
+        text=tokenizer.decode(all_tokens[len(initial_prompt_tokens) :]),
+        segments=all_segments,
+        language=language,
+    )
+
+
+def cli():
+    from . import available_models
+
+    def valid_model_name(name):
+        if name in available_models() or os.path.exists(name):
+            return name
+        raise ValueError(
+            f"model should be one of {available_models()} or path to a model checkpoint"
+        )
+
+    # fmt: off
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument("audio", nargs="+", type=str, help="audio file(s) to transcribe")
+    parser.add_argument("--model", default="turbo", type=valid_model_name, help="name of the Whisper model to use")
+    parser.add_argument("--model_dir", type=str, default=None, help="the path to save model files; uses ~/.cache/whisper by default")
+    parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu", help="device to use for PyTorch inference")
+    parser.add_argument("--output_dir", "-o", type=str, default=".", help="directory to save the outputs")
+    parser.add_argument("--output_format", "-f", type=str, default="all", choices=["txt", "vtt", "srt", "tsv", "json", "all"], help="format of the output file; if not specified, all available formats will be produced")
+    parser.add_argument("--verbose", type=str2bool, default=True, help="whether to print out the progress and debug messages")
+
+    parser.add_argument("--task", type=str, default="transcribe", choices=["transcribe", "translate"], help="whether to perform X->X speech recognition ('transcribe') or X->English translation ('translate')")
+    parser.add_argument("--language", type=str, default=None, choices=sorted(LANGUAGES.keys()) + sorted([k.title() for k in TO_LANGUAGE_CODE.keys()]), help="language spoken in the audio, specify None to perform language detection")
+
+    parser.add_argument("--temperature", type=float, default=0, help="temperature to use for sampling")
+    parser.add_argument("--best_of", type=optional_int, default=5, help="number of candidates when sampling with non-zero temperature")
+    parser.add_argument("--beam_size", type=optional_int, default=5, help="number of beams in beam search, only applicable when temperature is zero")
+    parser.add_argument("--patience", type=float, default=None, help="optional patience value to use in beam decoding, as in https://arxiv.org/abs/2204.05424, the default (1.0) is equivalent to conventional beam search")
+    parser.add_argument("--length_penalty", type=float, default=None, help="optional token length penalty coefficient (alpha) as in https://arxiv.org/abs/1609.08144, uses simple length normalization by default")
+
+    parser.add_argument("--suppress_tokens", type=str, default="-1", help="comma-separated list of token ids to suppress during sampling; '-1' will suppress most special characters except common punctuations")
+    parser.add_argument("--initial_prompt", type=str, default=None, help="optional text to provide as a prompt for the first window.")
+    parser.add_argument("--carry_initial_prompt", type=str2bool, default=False, help="if True, prepend initial_prompt to every internal decode() call. May reduce the effectiveness of condition_on_previous_text")
+
+    parser.add_argument("--condition_on_previous_text", type=str2bool, default=True, help="if True, provide the previous output of the model as a prompt for the next window; disabling may make the text inconsistent across windows, but the model becomes less prone to getting stuck in a failure loop")
+    parser.add_argument("--fp16", type=str2bool, default=True, help="whether to perform inference in fp16; True by default")
+
+    parser.add_argument("--temperature_increment_on_fallback", type=optional_float, default=0.2, help="temperature to increase when falling back when the decoding fails to meet either of the thresholds below")
+    parser.add_argument("--compression_ratio_threshold", type=optional_float, default=2.4, help="if the gzip compression ratio is higher than this value, treat the decoding as failed")
+    parser.add_argument("--logprob_threshold", type=optional_float, default=-1.0, help="if the average log probability is lower than this value, treat the decoding as failed")
+    parser.add_argument("--no_speech_threshold", type=optional_float, default=0.6, help="if the probability of the <|nospeech|> token is higher than this value AND the decoding has failed due to `logprob_threshold`, consider the segment as silence")
+    parser.add_argument("--word_timestamps", type=str2bool, default=False, help="(experimental) extract word-level timestamps and refine the results based on them")
+    parser.add_argument("--prepend_punctuations", type=str, default="\"\'“¿([{-", help="if word_timestamps is True, merge these punctuation symbols with the next word")
+    parser.add_argument("--append_punctuations", type=str, default="\"\'.。,，!！?？:：”)]}、", help="if word_timestamps is True, merge these punctuation symbols with the previous word")
+    parser.add_argument("--highlight_words", type=str2bool, default=False, help="(requires --word_timestamps True) underline each word as it is spoken in srt and vtt")
+    parser.add_argument("--max_line_width", type=optional_int, default=None, help="(requires --word_timestamps True) the maximum number of characters in a line before breaking the line")
+    parser.add_argument("--max_line_count", type=optional_int, default=None, help="(requires --word_timestamps True) the maximum number of lines in a segment")
+    parser.add_argument("--max_words_per_line", type=optional_int, default=None, help="(requires --word_timestamps True, no effect with --max_line_width) the maximum number of words in a segment")
+    parser.add_argument("--threads", type=optional_int, default=0, help="number of threads used by torch for CPU inference; supercedes MKL_NUM_THREADS/OMP_NUM_THREADS")
+    parser.add_argument("--clip_timestamps", type=str, default="0", help="comma-separated list start,end,start,end,... timestamps (in seconds) of clips to process, where the last end timestamp defaults to the end of the file")
+    parser.add_argument("--hallucination_silence_threshold", type=optional_float, help="(requires --word_timestamps True) skip silent periods longer than this threshold (in seconds) when a possible hallucination is detected")
+    # fmt: on
+
+    args = parser.parse_args().__dict__
+    model_name: str = args.pop("model")
+    model_dir: str = args.pop("model_dir")
+    output_dir: str = args.pop("output_dir")
+    output_format: str = args.pop("output_format")
+    device: str = args.pop("device")
+    os.makedirs(output_dir, exist_ok=True)
+
+    if model_name.endswith(".en") and args["language"] not in {"en", "English"}:
+        if args["language"] is not None:
+            warnings.warn(
+                f"{model_name} is an English-only model but receipted '{args['language']}'; using English instead."
+            )
+        args["language"] = "en"
+
+    temperature = args.pop("temperature")
+    if (increment := args.pop("temperature_increment_on_fallback")) is not None:
+        temperature = tuple(np.arange(temperature, 1.0 + 1e-6, increment))
+    else:
+        temperature = [temperature]
+
+    if (threads := args.pop("threads")) > 0:
+        torch.set_num_threads(threads)
+
+    from . import load_model
+
+    model = load_model(model_name, device=device, download_root=model_dir)
+
+    writer = get_writer(output_format, output_dir)
+    word_options = [
+        "highlight_words",
+        "max_line_count",
+        "max_line_width",
+        "max_words_per_line",
+    ]
+    if not args["word_timestamps"]:
+        for option in word_options:
+            if args[option]:
+                parser.error(f"--{option} requires --word_timestamps True")
+    if args["max_line_count"] and not args["max_line_width"]:
+        warnings.warn("--max_line_count has no effect without --max_line_width")
+    if args["max_words_per_line"] and args["max_line_width"]:
+        warnings.warn("--max_words_per_line has no effect with --max_line_width")
+    writer_args = {arg: args.pop(arg) for arg in word_options}
+    for audio_path in args.pop("audio"):
+        try:
+            result = transcribe(model, audio_path, temperature=temperature, **args)
+            writer(result, audio_path, **writer_args)
+        except Exception as e:
+            traceback.print_exc()
+            print(f"Skipping {audio_path} due to {type(e).__name__}: {str(e)}")
+
+
+if __name__ == "__main__":
+    cli()

whisperlivekit/simul_whisper/whisper/triton_ops.py

@@ -0,0 +1,117 @@
+from functools import lru_cache
+
+import numpy as np
+import torch
+
+try:
+    import triton
+    import triton.language as tl
+except ImportError:
+    raise RuntimeError("triton import failed; try `pip install --pre triton`")
+
+
+@triton.jit
+def dtw_kernel(
+    cost, trace, x, x_stride, cost_stride, trace_stride, N, M, BLOCK_SIZE: tl.constexpr
+):
+    offsets = tl.arange(0, BLOCK_SIZE)
+    mask = offsets < M
+
+    for k in range(1, N + M + 1):  # k = i + j
+        tl.debug_barrier()
+
+        p0 = cost + (k - 1) * cost_stride
+        p1 = cost + k * cost_stride
+        p2 = cost + k * cost_stride + 1
+
+        c0 = tl.load(p0 + offsets, mask=mask)
+        c1 = tl.load(p1 + offsets, mask=mask)
+        c2 = tl.load(p2 + offsets, mask=mask)
+
+        x_row = tl.load(x + (k - 1) * x_stride + offsets, mask=mask, other=0)
+        cost_row = x_row + tl.minimum(tl.minimum(c0, c1), c2)
+
+        cost_ptr = cost + (k + 1) * cost_stride + 1
+        tl.store(cost_ptr + offsets, cost_row, mask=mask)
+
+        trace_ptr = trace + (k + 1) * trace_stride + 1
+        tl.store(trace_ptr + offsets, 2, mask=mask & (c2 <= c0) & (c2 <= c1))
+        tl.store(trace_ptr + offsets, 1, mask=mask & (c1 <= c0) & (c1 <= c2))
+        tl.store(trace_ptr + offsets, 0, mask=mask & (c0 <= c1) & (c0 <= c2))
+
+
+@lru_cache(maxsize=None)
+def median_kernel(filter_width: int):
+    @triton.jit
+    def kernel(
+        y, x, x_stride, y_stride, BLOCK_SIZE: tl.constexpr
+    ):  # x.shape[-1] == filter_width
+        row_idx = tl.program_id(0)
+        offsets = tl.arange(0, BLOCK_SIZE)
+        mask = offsets < y_stride
+
+        x_ptr = x + row_idx * x_stride  # noqa: F841
+        y_ptr = y + row_idx * y_stride
+
+        LOAD_ALL_ROWS_HERE  # noqa: F821
+
+        BUBBLESORT_HERE  # noqa: F821
+
+        tl.store(y_ptr + offsets, MIDDLE_ROW_HERE, mask=mask)  # noqa: F821
+
+    kernel = triton.JITFunction(kernel.fn)
+    new_kernel = kernel.src.replace(
+        "    LOAD_ALL_ROWS_HERE",
+        "\n".join(
+            [
+                f"    row{i} = tl.load(x_ptr + offsets + {i}, mask=mask)"
+                for i in range(filter_width)
+            ]
+        ),
+    )
+
+    new_kernel = new_kernel.replace(
+        "    BUBBLESORT_HERE",
+        "\n\n".join(
+            [
+                "\n\n".join(
+                    [
+                        "\n".join(
+                            [
+                                f"    smaller = tl.where(row{j} < row{j + 1}, row{j}, row{j + 1})",
+                                f"    larger = tl.where(row{j} > row{j + 1}, row{j}, row{j + 1})",
+                                f"    row{j} = smaller",
+                                f"    row{j + 1} = larger",
+                            ]
+                        )
+                        for j in range(filter_width - i - 1)
+                    ]
+                )
+                for i in range(filter_width // 2 + 1)
+            ]
+        ),
+    )
+
+    new_kernel = new_kernel.replace("MIDDLE_ROW_HERE", f"row{filter_width // 2}")
+
+    if hasattr(kernel, "_unsafe_update_src") is True:
+        kernel._unsafe_update_src(new_kernel)
+        kernel.hash = None
+    else:
+        kernel.src = new_kernel
+
+    return kernel
+
+
+def median_filter_cuda(x: torch.Tensor, filter_width: int):
+    """Apply a median filter of given width along the last dimension of x"""
+    slices = x.contiguous().unfold(-1, filter_width, 1)
+    grid = np.prod(slices.shape[:-2])
+
+    kernel = median_kernel(filter_width)
+    y = torch.empty_like(slices[..., 0])
+
+    BLOCK_SIZE = 1 << (y.stride(-2) - 1).bit_length()
+    kernel[(grid,)](y, x, x.stride(-2), y.stride(-2), BLOCK_SIZE=BLOCK_SIZE)
+
+    return y

whisperlivekit/simul_whisper/whisper/utils.py

@@ -0,0 +1,318 @@
+import json
+import os
+import re
+import sys
+import zlib
+from typing import Callable, List, Optional, TextIO
+
+system_encoding = sys.getdefaultencoding()
+
+if system_encoding != "utf-8":
+
+    def make_safe(string):
+        # replaces any character not representable using the system default encoding with an '?',
+        # avoiding UnicodeEncodeError (https://github.com/openai/whisper/discussions/729).
+        return string.encode(system_encoding, errors="replace").decode(system_encoding)
+
+else:
+
+    def make_safe(string):
+        # utf-8 can encode any Unicode code point, so no need to do the round-trip encoding
+        return string
+
+
+def exact_div(x, y):
+    assert x % y == 0
+    return x // y
+
+
+def str2bool(string):
+    str2val = {"True": True, "False": False}
+    if string in str2val:
+        return str2val[string]
+    else:
+        raise ValueError(f"Expected one of {set(str2val.keys())}, got {string}")
+
+
+def optional_int(string):
+    return None if string == "None" else int(string)
+
+
+def optional_float(string):
+    return None if string == "None" else float(string)
+
+
+def compression_ratio(text) -> float:
+    text_bytes = text.encode("utf-8")
+    return len(text_bytes) / len(zlib.compress(text_bytes))
+
+
+def format_timestamp(
+    seconds: float, always_include_hours: bool = False, decimal_marker: str = "."
+):
+    assert seconds >= 0, "non-negative timestamp expected"
+    milliseconds = round(seconds * 1000.0)
+
+    hours = milliseconds // 3_600_000
+    milliseconds -= hours * 3_600_000
+
+    minutes = milliseconds // 60_000
+    milliseconds -= minutes * 60_000
+
+    seconds = milliseconds // 1_000
+    milliseconds -= seconds * 1_000
+
+    hours_marker = f"{hours:02d}:" if always_include_hours or hours > 0 else ""
+    return (
+        f"{hours_marker}{minutes:02d}:{seconds:02d}{decimal_marker}{milliseconds:03d}"
+    )
+
+
+def get_start(segments: List[dict]) -> Optional[float]:
+    return next(
+        (w["start"] for s in segments for w in s["words"]),
+        segments[0]["start"] if segments else None,
+    )
+
+
+def get_end(segments: List[dict]) -> Optional[float]:
+    return next(
+        (w["end"] for s in reversed(segments) for w in reversed(s["words"])),
+        segments[-1]["end"] if segments else None,
+    )
+
+
+class ResultWriter:
+    extension: str
+
+    def __init__(self, output_dir: str):
+        self.output_dir = output_dir
+
+    def __call__(
+        self, result: dict, audio_path: str, options: Optional[dict] = None, **kwargs
+    ):
+        audio_basename = os.path.basename(audio_path)
+        audio_basename = os.path.splitext(audio_basename)[0]
+        output_path = os.path.join(
+            self.output_dir, audio_basename + "." + self.extension
+        )
+
+        with open(output_path, "w", encoding="utf-8") as f:
+            self.write_result(result, file=f, options=options, **kwargs)
+
+    def write_result(
+        self, result: dict, file: TextIO, options: Optional[dict] = None, **kwargs
+    ):
+        raise NotImplementedError
+
+
+class WriteTXT(ResultWriter):
+    extension: str = "txt"
+
+    def write_result(
+        self, result: dict, file: TextIO, options: Optional[dict] = None, **kwargs
+    ):
+        for segment in result["segments"]:
+            print(segment["text"].strip(), file=file, flush=True)
+
+
+class SubtitlesWriter(ResultWriter):
+    always_include_hours: bool
+    decimal_marker: str
+
+    def iterate_result(
+        self,
+        result: dict,
+        options: Optional[dict] = None,
+        *,
+        max_line_width: Optional[int] = None,
+        max_line_count: Optional[int] = None,
+        highlight_words: bool = False,
+        max_words_per_line: Optional[int] = None,
+    ):
+        options = options or {}
+        max_line_width = max_line_width or options.get("max_line_width")
+        max_line_count = max_line_count or options.get("max_line_count")
+        highlight_words = highlight_words or options.get("highlight_words", False)
+        max_words_per_line = max_words_per_line or options.get("max_words_per_line")
+        preserve_segments = max_line_count is None or max_line_width is None
+        max_line_width = max_line_width or 1000
+        max_words_per_line = max_words_per_line or 1000
+
+        def iterate_subtitles():
+            line_len = 0
+            line_count = 1
+            # the next subtitle to yield (a list of word timings with whitespace)
+            subtitle: List[dict] = []
+            last: float = get_start(result["segments"]) or 0.0
+            for segment in result["segments"]:
+                chunk_index = 0
+                words_count = max_words_per_line
+                while chunk_index < len(segment["words"]):
+                    remaining_words = len(segment["words"]) - chunk_index
+                    if max_words_per_line > len(segment["words"]) - chunk_index:
+                        words_count = remaining_words
+                    for i, original_timing in enumerate(
+                        segment["words"][chunk_index : chunk_index + words_count]
+                    ):
+                        timing = original_timing.copy()
+                        long_pause = (
+                            not preserve_segments and timing["start"] - last > 3.0
+                        )
+                        has_room = line_len + len(timing["word"]) <= max_line_width
+                        seg_break = i == 0 and len(subtitle) > 0 and preserve_segments
+                        if (
+                            line_len > 0
+                            and has_room
+                            and not long_pause
+                            and not seg_break
+                        ):
+                            # line continuation
+                            line_len += len(timing["word"])
+                        else:
+                            # new line
+                            timing["word"] = timing["word"].strip()
+                            if (
+                                len(subtitle) > 0
+                                and max_line_count is not None
+                                and (long_pause or line_count >= max_line_count)
+                                or seg_break
+                            ):
+                                # subtitle break
+                                yield subtitle
+                                subtitle = []
+                                line_count = 1
+                            elif line_len > 0:
+                                # line break
+                                line_count += 1
+                                timing["word"] = "\n" + timing["word"]
+                            line_len = len(timing["word"].strip())
+                        subtitle.append(timing)
+                        last = timing["start"]
+                    chunk_index += max_words_per_line
+            if len(subtitle) > 0:
+                yield subtitle
+
+        if len(result["segments"]) > 0 and "words" in result["segments"][0]:
+            for subtitle in iterate_subtitles():
+                subtitle_start = self.format_timestamp(subtitle[0]["start"])
+                subtitle_end = self.format_timestamp(subtitle[-1]["end"])
+                subtitle_text = "".join([word["word"] for word in subtitle])
+                if highlight_words:
+                    last = subtitle_start
+                    all_words = [timing["word"] for timing in subtitle]
+                    for i, this_word in enumerate(subtitle):
+                        start = self.format_timestamp(this_word["start"])
+                        end = self.format_timestamp(this_word["end"])
+                        if last != start:
+                            yield last, start, subtitle_text
+
+                        yield start, end, "".join(
+                            [
+                                (
+                                    re.sub(r"^(\s*)(.*)$", r"\1<u>\2</u>", word)
+                                    if j == i
+                                    else word
+                                )
+                                for j, word in enumerate(all_words)
+                            ]
+                        )
+                        last = end
+                else:
+                    yield subtitle_start, subtitle_end, subtitle_text
+        else:
+            for segment in result["segments"]:
+                segment_start = self.format_timestamp(segment["start"])
+                segment_end = self.format_timestamp(segment["end"])
+                segment_text = segment["text"].strip().replace("-->", "->")
+                yield segment_start, segment_end, segment_text
+
+    def format_timestamp(self, seconds: float):
+        return format_timestamp(
+            seconds=seconds,
+            always_include_hours=self.always_include_hours,
+            decimal_marker=self.decimal_marker,
+        )
+
+
+class WriteVTT(SubtitlesWriter):
+    extension: str = "vtt"
+    always_include_hours: bool = False
+    decimal_marker: str = "."
+
+    def write_result(
+        self, result: dict, file: TextIO, options: Optional[dict] = None, **kwargs
+    ):
+        print("WEBVTT\n", file=file)
+        for start, end, text in self.iterate_result(result, options, **kwargs):
+            print(f"{start} --> {end}\n{text}\n", file=file, flush=True)
+
+
+class WriteSRT(SubtitlesWriter):
+    extension: str = "srt"
+    always_include_hours: bool = True
+    decimal_marker: str = ","
+
+    def write_result(
+        self, result: dict, file: TextIO, options: Optional[dict] = None, **kwargs
+    ):
+        for i, (start, end, text) in enumerate(
+            self.iterate_result(result, options, **kwargs), start=1
+        ):
+            print(f"{i}\n{start} --> {end}\n{text}\n", file=file, flush=True)
+
+
+class WriteTSV(ResultWriter):
+    """
+    Write a transcript to a file in TSV (tab-separated values) format containing lines like:
+    <start time in integer milliseconds>\t<end time in integer milliseconds>\t<transcript text>
+
+    Using integer milliseconds as start and end times means there's no chance of interference from
+    an environment setting a language encoding that causes the decimal in a floating point number
+    to appear as a comma; also is faster and more efficient to parse & store, e.g., in C++.
+    """
+
+    extension: str = "tsv"
+
+    def write_result(
+        self, result: dict, file: TextIO, options: Optional[dict] = None, **kwargs
+    ):
+        print("start", "end", "text", sep="\t", file=file)
+        for segment in result["segments"]:
+            print(round(1000 * segment["start"]), file=file, end="\t")
+            print(round(1000 * segment["end"]), file=file, end="\t")
+            print(segment["text"].strip().replace("\t", " "), file=file, flush=True)
+
+
+class WriteJSON(ResultWriter):
+    extension: str = "json"
+
+    def write_result(
+        self, result: dict, file: TextIO, options: Optional[dict] = None, **kwargs
+    ):
+        json.dump(result, file)
+
+
+def get_writer(
+    output_format: str, output_dir: str
+) -> Callable[[dict, TextIO, dict], None]:
+    writers = {
+        "txt": WriteTXT,
+        "vtt": WriteVTT,
+        "srt": WriteSRT,
+        "tsv": WriteTSV,
+        "json": WriteJSON,
+    }
+
+    if output_format == "all":
+        all_writers = [writer(output_dir) for writer in writers.values()]
+
+        def write_all(
+            result: dict, file: TextIO, options: Optional[dict] = None, **kwargs
+        ):
+            for writer in all_writers:
+                writer(result, file, options, **kwargs)
+
+        return write_all
+
+    return writers[output_format](output_dir)

whisperlivekit/simul_whisper/whisper/version.py

@@ -0,0 +1 @@
+__version__ = "20250625"

whisperlivekit/timed_objects.py

@@ -29,4 +29,8 @@ class SpeakerSegment(TimedText):
     """Represents a segment of audio attributed to a specific speaker.
     No text nor probability is associated with this segment.
     """
-    pass
+    pass
+
+@dataclass
+class Silence():
+    duration: float

whisperlivekit/trail_repetition.py

@@ -0,0 +1,60 @@
+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/warmup.py

@@ -0,0 +1,62 @@
+
+import logging
+
+logger = logging.getLogger(__name__)
+
+def load_file(warmup_file=None, timeout=5):
+    import os
+    import tempfile
+    import librosa
+        
+    if warmup_file is None:
+        # Download JFK sample if not already present
+        jfk_url = "https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav"
+        temp_dir = tempfile.gettempdir()
+        warmup_file = os.path.join(temp_dir, "whisper_warmup_jfk.wav")
+        
+        if not os.path.exists(warmup_file):
+            logger.debug(f"Downloading warmup file from {jfk_url}")
+            print(f"Downloading warmup file from {jfk_url}")
+            import time
+            import urllib.request
+            import urllib.error
+            import socket
+            
+            original_timeout = socket.getdefaulttimeout()
+            socket.setdefaulttimeout(timeout)
+            
+            start_time = time.time()
+            try:
+                urllib.request.urlretrieve(jfk_url, warmup_file)
+                logger.debug(f"Download successful in {time.time() - start_time:.2f}s")
+            except (urllib.error.URLError, socket.timeout) as e:
+                logger.warning(f"Download failed: {e}. Proceeding without warmup.")
+                return False
+            finally:
+                socket.setdefaulttimeout(original_timeout)
+    elif not warmup_file:
+        return False 
+    
+    if not warmup_file or not os.path.exists(warmup_file) or os.path.getsize(warmup_file) == 0:
+        logger.warning(f"Warmup file {warmup_file} invalid or missing.")
+        return False
+    
+    try:
+        audio, sr = librosa.load(warmup_file, sr=16000)
+    except Exception as e:
+        logger.warning(f"Failed to load audio file: {e}")
+        return False
+    return audio
+
+def warmup_asr(asr, warmup_file=None, timeout=5):
+    """
+    Warmup the ASR model by transcribing a short audio file.
+    """
+    audio = load_file(warmup_file=None, timeout=5)
+    asr.transcribe(audio)
+    logger.info("ASR model is warmed up")
+    
+def warmup_online(online, warmup_file=None, timeout=5):
+    audio = load_file(warmup_file=None, timeout=5)
+    online.warmup(audio)
+    logger.warning("ASR is warmed up")

whisperlivekit/web/live_transcription.css

@@ -0,0 +1,402 @@
+:root {
+  --bg: #ffffff;
+  --text: #111111;
+  --muted: #666666;
+  --border: #e5e5e5;
+  --chip-bg: rgba(0, 0, 0, 0.04);
+  --chip-text: #000000;
+  --spinner-border: #8d8d8d5c;
+  --spinner-top: #b0b0b0;
+  --silence-bg: #f3f3f3;
+  --loading-bg: rgba(255, 77, 77, 0.06);
+  --button-bg: #ffffff;
+  --button-border: #e9e9e9;
+  --wave-stroke: #000000;
+  --label-dia-text: #868686;
+  --label-trans-text: #111111;
+}
+
+@media (prefers-color-scheme: dark) {
+  :root:not([data-theme="light"]) {
+    --bg: #0b0b0b;
+    --text: #e6e6e6;
+    --muted: #9aa0a6;
+    --border: #333333;
+    --chip-bg: rgba(255, 255, 255, 0.08);
+    --chip-text: #e6e6e6;
+    --spinner-border: #555555;
+    --spinner-top: #dddddd;
+    --silence-bg: #1a1a1a;
+    --loading-bg: rgba(255, 77, 77, 0.12);
+    --button-bg: #111111;
+    --button-border: #333333;
+    --wave-stroke: #e6e6e6;
+    --label-dia-text: #b3b3b3;
+    --label-trans-text: #ffffff;
+  }
+}
+
+:root[data-theme="dark"] {
+  --bg: #0b0b0b;
+  --text: #e6e6e6;
+  --muted: #9aa0a6;
+  --border: #333333;
+  --chip-bg: rgba(255, 255, 255, 0.08);
+  --chip-text: #e6e6e6;
+  --spinner-border: #555555;
+  --spinner-top: #dddddd;
+  --silence-bg: #1a1a1a;
+  --loading-bg: rgba(255, 77, 77, 0.12);
+  --button-bg: #111111;
+  --button-border: #333333;
+  --wave-stroke: #e6e6e6;
+  --label-dia-text: #b3b3b3;
+  --label-trans-text: #ffffff;
+}
+
+:root[data-theme="light"] {
+  --bg: #ffffff;
+  --text: #111111;
+  --muted: #666666;
+  --border: #e5e5e5;
+  --chip-bg: rgba(0, 0, 0, 0.04);
+  --chip-text: #000000;
+  --spinner-border: #8d8d8d5c;
+  --spinner-top: #b0b0b0;
+  --silence-bg: #f3f3f3;
+  --loading-bg: rgba(255, 77, 77, 0.06);
+  --button-bg: #ffffff;
+  --button-border: #e9e9e9;
+  --wave-stroke: #000000;
+  --label-dia-text: #868686;
+  --label-trans-text: #111111;
+}
+
+body {
+  font-family: ui-sans-serif, system-ui, sans-serif, 'Apple Color Emoji', 'Segoe UI Emoji', 'Segoe UI Symbol', 'Noto Color Emoji';
+  margin: 20px;
+  text-align: center;
+  background-color: var(--bg);
+  color: var(--text);
+}
+
+/* Record button */
+#recordButton {
+  width: 50px;
+  height: 50px;
+  border: none;
+  border-radius: 50%;
+  background-color: var(--button-bg);
+  cursor: pointer;
+  transition: all 0.3s ease;
+  border: 1px solid var(--button-border);
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  position: relative;
+}
+
+#recordButton.recording {
+  width: 180px;
+  border-radius: 40px;
+  justify-content: flex-start;
+  padding-left: 20px;
+}
+
+#recordButton:active {
+  transform: scale(0.95);
+}
+
+.shape-container {
+  width: 25px;
+  height: 25px;
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  flex-shrink: 0;
+}
+
+.shape {
+  width: 25px;
+  height: 25px;
+  background-color: rgb(209, 61, 53);
+  border-radius: 50%;
+  transition: all 0.3s ease;
+}
+
+#recordButton:disabled .shape {
+  background-color: #6e6d6d;
+}
+
+#recordButton.recording .shape {
+  border-radius: 5px;
+  width: 25px;
+  height: 25px;
+}
+
+/* Recording elements */
+.recording-info {
+  display: none;
+  align-items: center;
+  margin-left: 15px;
+  flex-grow: 1;
+}
+
+#recordButton.recording .recording-info {
+  display: flex;
+}
+
+.wave-container {
+  width: 60px;
+  height: 30px;
+  position: relative;
+  display: flex;
+  align-items: center;
+  justify-content: center;
+}
+
+#waveCanvas {
+  width: 100%;
+  height: 100%;
+}
+
+.timer {
+  font-size: 14px;
+  font-weight: 500;
+  color: var(--text);
+  margin-left: 10px;
+}
+
+#status {
+  margin-top: 20px;
+  font-size: 16px;
+  color: var(--text);
+}
+
+/* Settings */
+.settings-container {
+  display: flex;
+  justify-content: center;
+  align-items: center;
+  gap: 15px;
+  margin-top: 20px;
+}
+
+.settings {
+  display: flex;
+  flex-direction: column;
+  align-items: flex-start;
+  gap: 12px;
+}
+
+.field {
+  display: flex;
+  flex-direction: column;
+  align-items: flex-start;
+  gap: 3px;
+}
+
+#chunkSelector,
+#websocketInput,
+#themeSelector {
+  font-size: 16px;
+  padding: 5px 8px;
+  border-radius: 8px;
+  border: 1px solid var(--border);
+  background-color: var(--button-bg);
+  color: var(--text);
+  max-height: 34px;
+}
+
+#websocketInput {
+  width: 220px;
+}
+
+#chunkSelector:focus,
+#websocketInput:focus,
+#themeSelector:focus {
+  outline: none;
+  border-color: #007bff;
+  box-shadow: 0 0 0 3px rgba(0, 123, 255, 0.15);
+}
+
+label {
+  font-size: 13px;
+  color: var(--muted);
+}
+
+.ws-default {
+  font-size: 12px;
+  color: var(--muted);
+}
+
+/* Segmented pill control for Theme */
+.segmented {
+  display: inline-flex;
+  align-items: stretch;
+  border: 1px solid var(--button-border);
+  background-color: var(--button-bg);
+  border-radius: 999px;
+  overflow: hidden;
+}
+
+.segmented input[type="radio"] {
+  position: absolute;
+  opacity: 0;
+  pointer-events: none;
+}
+
+.theme-selector-container {
+  position: absolute;
+  top: 20px;
+  right: 20px;
+}
+
+.segmented label {
+  display: inline-flex;
+  align-items: center;
+  gap: 6px;
+  padding: 6px 12px;
+  font-size: 14px;
+  color: var(--muted);
+  cursor: pointer;
+  user-select: none;
+  transition: background-color 0.2s ease, color 0.2s ease;
+}
+
+.segmented label span {
+  display: none;
+}
+
+.segmented label:hover span {
+  display: inline;
+}
+
+.segmented label:hover {
+  background-color: var(--chip-bg);
+}
+
+.segmented img {
+  width: 16px;
+  height: 16px;
+}
+
+.segmented input[type="radio"]:checked + label {
+  background-color: var(--chip-bg);
+  color: var(--text);
+}
+
+.segmented input[type="radio"]:focus-visible + label,
+.segmented input[type="radio"]:focus + label {
+  outline: 2px solid #007bff;
+  outline-offset: 2px;
+  border-radius: 999px;
+}
+
+/* Transcript area */
+#linesTranscript {
+  margin: 20px auto;
+  max-width: 700px;
+  text-align: left;
+  font-size: 16px;
+}
+
+#linesTranscript p {
+  margin: 0px 0;
+}
+
+#linesTranscript strong {
+  color: var(--text);
+}
+
+#speaker {
+  border: 1px solid var(--border);
+  border-radius: 100px;
+  padding: 2px 10px;
+  font-size: 14px;
+  margin-bottom: 0px;
+}
+
+.label_diarization {
+  background-color: var(--chip-bg);
+  border-radius: 8px 8px 8px 8px;
+  padding: 2px 10px;
+  margin-left: 10px;
+  display: inline-block;
+  white-space: nowrap;
+  font-size: 14px;
+  margin-bottom: 0px;
+  color: var(--label-dia-text);
+}
+
+.label_transcription {
+  background-color: var(--chip-bg);
+  border-radius: 8px 8px 8px 8px;
+  padding: 2px 10px;
+  display: inline-block;
+  white-space: nowrap;
+  margin-left: 10px;
+  font-size: 14px;
+  margin-bottom: 0px;
+  color: var(--label-trans-text);
+}
+
+#timeInfo {
+  color: var(--muted);
+  margin-left: 10px;
+}
+
+.textcontent {
+  font-size: 16px;
+  padding-left: 10px;
+  margin-bottom: 10px;
+  margin-top: 1px;
+  padding-top: 5px;
+  border-radius: 0px 0px 0px 10px;
+}
+
+.buffer_diarization {
+  color: var(--label-dia-text);
+  margin-left: 4px;
+}
+
+.buffer_transcription {
+  color: #7474748c;
+  margin-left: 4px;
+}
+
+.spinner {
+  display: inline-block;
+  width: 8px;
+  height: 8px;
+  border: 2px solid var(--spinner-border);
+  border-top: 2px solid var(--spinner-top);
+  border-radius: 50%;
+  animation: spin 0.7s linear infinite;
+  vertical-align: middle;
+  margin-bottom: 2px;
+  margin-right: 5px;
+}
+
+@keyframes spin {
+  to {
+    transform: rotate(360deg);
+  }
+}
+
+.silence {
+  color: var(--muted);
+  background-color: var(--silence-bg);
+  font-size: 13px;
+  border-radius: 30px;
+  padding: 2px 10px;
+}
+
+.loading {
+  color: var(--muted);
+  background-color: var(--loading-bg);
+  border-radius: 8px 8px 8px 0px;
+  padding: 2px 10px;
+  font-size: 14px;
+  margin-bottom: 0px;
+}

whisperlivekit/web/live_transcription.html

@@ -1,682 +1,61 @@
 <!DOCTYPE html>
 <html lang="en">
-
 <head>
-    <meta charset="UTF-8" />
-    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
-    <title>Audio Transcription</title>
-    <style>
-        body {
-            font-family: ui-sans-serif, system-ui, sans-serif, 'Apple Color Emoji', 'Segoe UI Emoji', 'Segoe UI Symbol', 'Noto Color Emoji';
-            margin: 20px;
-            text-align: center;
-        }
-
-        #recordButton {
-            width: 50px;
-            height: 50px;
-            border: none;
-            border-radius: 50%;
-            background-color: white;
-            cursor: pointer;
-            transition: all 0.3s ease;
-            border: 1px solid rgb(233, 233, 233);
-            display: flex;
-            align-items: center;
-            justify-content: center;
-            position: relative;
-        }
-
-        #recordButton.recording {
-            width: 180px;
-            border-radius: 40px;
-            justify-content: flex-start;
-            padding-left: 20px;
-        }
-
-        #recordButton:active {
-            transform: scale(0.95);
-        }
-
-        .shape-container {
-            width: 25px;
-            height: 25px;
-            display: flex;
-            align-items: center;
-            justify-content: center;
-            flex-shrink: 0;
-        }
-
-        .shape {
-            width: 25px;
-            height: 25px;
-            background-color: rgb(209, 61, 53);
-            border-radius: 50%;
-            transition: all 0.3s ease;
-        }
-
-        #recordButton:disabled .shape {
-            background-color: #6e6d6d;
-        }
-
-        #recordButton.recording .shape {
-            border-radius: 5px;
-            width: 25px;
-            height: 25px;
-        }
-
-        /* Recording elements */
-        .recording-info {
-            display: none;
-            align-items: center;
-            margin-left: 15px;
-            flex-grow: 1;
-        }
-
-        #recordButton.recording .recording-info {
-            display: flex;
-        }
-
-        .wave-container {
-            width: 60px;
-            height: 30px;
-            position: relative;
-            display: flex;
-            align-items: center;
-            justify-content: center;
-        }
-
-        #waveCanvas {
-            width: 100%;
-            height: 100%;
-        }
-
-        .timer {
-            font-size: 14px;
-            font-weight: 500;
-            color: #333;
-            margin-left: 10px;
-        }
-
-        #status {
-            margin-top: 20px;
-            font-size: 16px;
-            color: #333;
-        }
-
-        .settings-container {
-            display: flex;
-            justify-content: center;
-            align-items: center;
-            gap: 15px;
-            margin-top: 20px;
-        }
-
-        .settings {
-            display: flex;
-            flex-direction: column;
-            align-items: flex-start;
-            gap: 5px;
-        }
-
-        #chunkSelector,
-        #websocketInput {
-            font-size: 16px;
-            padding: 5px;
-            border-radius: 5px;
-            border: 1px solid #ddd;
-            background-color: #ffffff;
-            max-height: 30px;
-        }
-
-        #websocketInput {
-            width: 200px;
-        }
-
-        #chunkSelector:focus,
-        #websocketInput:focus {
-            outline: none;
-            border-color: #007bff;
-        }
-
-        label {
-            font-size: 14px;
-        }
-
-        /* Speaker-labeled transcript area */
-        #linesTranscript {
-            margin: 20px auto;
-            max-width: 700px;
-            text-align: left;
-            font-size: 16px;
-        }
-
-        #linesTranscript p {
-            margin: 0px 0;
-        }
-
-        #linesTranscript strong {
-            color: #333;
-        }
-
-        #speaker {
-            border: 1px solid rgb(229, 229, 229);
-            border-radius: 100px;
-            padding: 2px 10px;
-            font-size: 14px;
-            margin-bottom: 0px;
-        }
-        .label_diarization {
-            background-color: #ffffff66;
-            border-radius: 8px 8px 8px 8px;
-            padding: 2px 10px;
-            margin-left: 10px;
-            display: inline-block;
-            white-space: nowrap;
-            font-size: 14px;
-            margin-bottom: 0px;
-            color: rgb(134, 134, 134)
-        }
-
-        .label_transcription {
-            background-color: #ffffff66;
-            border-radius: 8px 8px 8px 8px;
-            padding: 2px 10px;
-            display: inline-block;
-            white-space: nowrap;
-            margin-left: 10px;
-            font-size: 14px;
-            margin-bottom: 0px;
-            color: #000000
-        }
-
-        #timeInfo {
-            color: #666;
-            margin-left: 10px;
-        }
-
-        .textcontent {
-            font-size: 16px;
-            /* margin-left: 10px; */
-            padding-left: 10px;
-            margin-bottom: 10px;
-            margin-top: 1px;
-            padding-top: 5px;
-            border-radius: 0px 0px 0px 10px;
-        }
-
-        .buffer_diarization {
-            color: rgb(134, 134, 134);
-            margin-left: 4px;
-        }
-
-        .buffer_transcription {
-            color: #7474748c;
-            margin-left: 4px;
-        }
-
-
-        .spinner {
-            display: inline-block;
-            width: 8px;
-            height: 8px;
-            border: 2px solid #8d8d8d5c;
-            border-top: 2px solid #6c6c6ce5;
-            border-radius: 50%;
-            animation: spin 0.6s linear infinite;
-            vertical-align: middle;
-            margin-bottom: 2px;
-            margin-right: 5px;
-        }
-
-        @keyframes spin {
-            to {
-                transform: rotate(360deg);
-            }
-        }
-
-        .silence {
-            color: #666;
-            background-color: #f3f3f3;
-            font-size: 13px;
-            border-radius: 30px;
-            padding: 2px 10px;
-        }
-
-        .loading {
-            color: #666;
-            background-color: #ff4d4d0f;
-            border-radius: 8px 8px 8px 0px;
-            padding: 2px 10px;
-            font-size: 14px;
-            margin-bottom: 0px;
-        }
-    </style>
+  <meta charset="UTF-8" />
+  <meta name="viewport" content="width=device-width, initial-scale=1.0" />
+  <title>WhisperLiveKit</title>
+  <link rel="stylesheet" href="/web/live_transcription.css" />
 </head>
-
 <body>
-
-    <div class="settings-container">
-        <button id="recordButton">
-            <div class="shape-container">
-                <div class="shape"></div>
-            </div>
-            <div class="recording-info">
-                <div class="wave-container">
-                    <canvas id="waveCanvas"></canvas>
-                </div>
-                <div class="timer">00:00</div>
-            </div>
-        </button>
-        <div class="settings">
-            <div>
-                <label for="chunkSelector">Chunk size (ms):</label>
-                <select id="chunkSelector">
-                    <option value="500">500 ms</option>
-                    <option value="1000" selected>1000 ms</option>
-                    <option value="2000">2000 ms</option>
-                    <option value="3000">3000 ms</option>
-                    <option value="4000">4000 ms</option>
-                    <option value="5000">5000 ms</option>
-                </select>
-            </div>
-            <div>
-                <label for="websocketInput">WebSocket URL:</label>
-                <input id="websocketInput" type="text" />
-            </div>
+  <div class="settings-container">
+    <button id="recordButton">
+      <div class="shape-container">
+        <div class="shape"></div>
+      </div>
+      <div class="recording-info">
+        <div class="wave-container">
+          <canvas id="waveCanvas"></canvas>
         </div>
+        <div class="timer">00:00</div>
+      </div>
+    </button>
+
+    <div class="settings">
+      <div class="field">
+        <label for="websocketInput">WebSocket URL</label>
+        <input id="websocketInput" type="text" placeholder="ws://host:port/asr" />
+      </div>
+
+      </div>
     </div>
+  </div>
 
-    <p id="status"></p>
+  <div class="theme-selector-container">
+    <div class="segmented" role="radiogroup" aria-label="Theme selector">
+      <input type="radio" id="theme-system" name="theme" value="system" />
+      <label for="theme-system" title="System">
+        <img src="/web/src/system_mode.svg" alt="" />
+        <span>System</span>
+      </label>
 
-    <!-- Speaker-labeled transcript -->
-    <div id="linesTranscript"></div>
+      <input type="radio" id="theme-light" name="theme" value="light" />
+      <label for="theme-light" title="Light">
+        <img src="/web/src/light_mode.svg" alt="" />
+        <span>Light</span>
+      </label>
 
-    <script>
-        let isRecording = false;
-        let websocket = null;
-        let recorder = null;
-        let chunkDuration = 1000;
-        let websocketUrl = "ws://localhost:8000/asr";
-        let userClosing = false;
-        let startTime = null;
-        let timerInterval = null;
-        let audioContext = null;
-        let analyser = null;
-        let microphone = null;
-        let waveCanvas = document.getElementById("waveCanvas");
-        let waveCtx = waveCanvas.getContext("2d");
-        let animationFrame = null;
-        let waitingForStop = false;
-        let lastReceivedData = null;
-        waveCanvas.width = 60 * (window.devicePixelRatio || 1);
-        waveCanvas.height = 30 * (window.devicePixelRatio || 1);
-        waveCtx.scale(window.devicePixelRatio || 1, window.devicePixelRatio || 1);
+      <input type="radio" id="theme-dark" name="theme" value="dark" />
+      <label for="theme-dark" title="Dark">
+        <img src="/web/src/dark_mode.svg" alt="" />
+        <span>Dark</span>
+      </label>
+    </div>
+  </div>
 
-        const statusText = document.getElementById("status");
-        const recordButton = document.getElementById("recordButton");
-        const chunkSelector = document.getElementById("chunkSelector");
-        const websocketInput = document.getElementById("websocketInput");
-        const linesTranscriptDiv = document.getElementById("linesTranscript");
-        const timerElement = document.querySelector(".timer");
+  <p id="status"></p>
 
-        const host = window.location.hostname || "localhost";
-        const port = window.location.port || "8000";
-        const protocol = window.location.protocol === "https:" ? "wss" : "ws";
-        const defaultWebSocketUrl = `${protocol}://${host}:${port}/asr`;
-        websocketInput.value = defaultWebSocketUrl;
-        websocketUrl = defaultWebSocketUrl;
+  <div id="linesTranscript"></div>
 
-        chunkSelector.addEventListener("change", () => {
-            chunkDuration = parseInt(chunkSelector.value);
-        });
-
-        websocketInput.addEventListener("change", () => {
-            const urlValue = websocketInput.value.trim();
-            if (!urlValue.startsWith("ws://") && !urlValue.startsWith("wss://")) {
-                statusText.textContent = "Invalid WebSocket URL (must start with ws:// or wss://)";
-                return;
-            }
-            websocketUrl = urlValue;
-            statusText.textContent = "WebSocket URL updated. Ready to connect.";
-        });
-
-        function setupWebSocket() {
-            return new Promise((resolve, reject) => {
-                try {
-                    websocket = new WebSocket(websocketUrl);
-                } catch (error) {
-                    statusText.textContent = "Invalid WebSocket URL. Please check and try again.";
-                    reject(error);
-                    return;
-                }
-
-                websocket.onopen = () => {
-                    statusText.textContent = "Connected to server.";
-                    resolve();
-                };
-
-                websocket.onclose = () => {
-                    if (userClosing) {
-                        if (waitingForStop) {
-                            statusText.textContent = "Processing finalized or connection closed.";
-                            if (lastReceivedData) {
-                                renderLinesWithBuffer(
-                                    lastReceivedData.lines || [],
-                                    lastReceivedData.buffer_diarization || "",
-                                    lastReceivedData.buffer_transcription || "",
-                                    0, 0, true // isFinalizing = true
-                                );
-                            }
-                        }
-                        // If ready_to_stop was received, statusText is already "Finished processing..."
-                        // and waitingForStop is false.
-                    } else {
-                        statusText.textContent = "Disconnected from the WebSocket server. (Check logs if model is loading.)";
-                        if (isRecording) {
-                            stopRecording(); 
-                        }
-                    }
-                    isRecording = false;  
-                    waitingForStop = false; 
-                    userClosing = false;  
-                    lastReceivedData = null;  
-                    websocket = null;    
-                    updateUI();  
-                };
-
-                websocket.onerror = () => {
-                    statusText.textContent = "Error connecting to WebSocket.";
-                    reject(new Error("Error connecting to WebSocket"));
-                };
-
-                // Handle messages from server
-                websocket.onmessage = (event) => {
-                    const data = JSON.parse(event.data);
-                    
-                    // Check for status messages
-                    if (data.type === "ready_to_stop") {
-                        console.log("Ready to stop received, finalizing display and closing WebSocket.");
-                        waitingForStop = false;
-
-                        if (lastReceivedData) {
-                            renderLinesWithBuffer(
-                                lastReceivedData.lines || [],
-                                lastReceivedData.buffer_diarization || "",
-                                lastReceivedData.buffer_transcription || "",
-                                0, // No more lag
-                                0, // No more lag
-                                true // isFinalizing = true
-                            );
-                        }
-                        statusText.textContent = "Finished processing audio! Ready to record again.";
-                        recordButton.disabled = false;
-                        
-                        if (websocket) {
-                            websocket.close(); // will trigger onclose
-                            // websocket = null; // onclose handle setting websocket to null
-                        }
-                        return;
-                    }
-                    
-                    lastReceivedData = data; 
-                    
-                    // Handle normal transcription updates
-                    const { 
-                        lines = [], 
-                        buffer_transcription = "", 
-                        buffer_diarization = "",
-                        remaining_time_transcription = 0,
-                        remaining_time_diarization = 0,
-                        status = "active_transcription"
-                    } = data;
-                    
-                    renderLinesWithBuffer(
-                        lines, 
-                        buffer_diarization, 
-                        buffer_transcription, 
-                        remaining_time_diarization,
-                        remaining_time_transcription,
-                        false,
-                        status
-                    );
-                };
-            });
-        }
-
-        function renderLinesWithBuffer(lines, buffer_diarization, buffer_transcription, remaining_time_diarization, remaining_time_transcription, isFinalizing = false, current_status = "active_transcription") {
-            if (current_status === "no_audio_detected") {
-                linesTranscriptDiv.innerHTML = "<p style='text-align: center; color: #666; margin-top: 20px;'><em>No audio detected...</em></p>";
-                return; 
-            }
-
-            const linesHtml = lines.map((item, idx) => {
-                let timeInfo = "";
-                if (item.beg !== undefined && item.end !== undefined) {
-                    timeInfo = ` ${item.beg} - ${item.end}`;
-                }
-
-                let speakerLabel = "";
-                if (item.speaker === -2) {
-                    speakerLabel = `<span class="silence">Silence<span id='timeInfo'>${timeInfo}</span></span>`;
-                } else if (item.speaker == 0 && !isFinalizing) {
-                    speakerLabel = `<span class='loading'><span class="spinner"></span><span id='timeInfo'>${remaining_time_diarization} second(s) of audio are undergoing diarization</span></span>`;
-                } else if (item.speaker == -1) {
-                    speakerLabel = `<span id="speaker">Speaker 1<span id='timeInfo'>${timeInfo}</span></span>`;
-                } else if (item.speaker !== -1 && item.speaker !== 0) {
-                    speakerLabel = `<span id="speaker">Speaker ${item.speaker}<span id='timeInfo'>${timeInfo}</span></span>`;
-                }
-
-
-                let currentLineText = item.text || "";
-
-                if (idx === lines.length - 1) { 
-                    if (!isFinalizing) {
-                        if (remaining_time_transcription > 0) {
-                             speakerLabel += `<span class="label_transcription"><span class="spinner"></span>Transcription lag <span id='timeInfo'>${remaining_time_transcription}s</span></span>`;
-                        }
-                        if (buffer_diarization && remaining_time_diarization > 0) {
-                             speakerLabel += `<span class="label_diarization"><span class="spinner"></span>Diarization lag<span id='timeInfo'>${remaining_time_diarization}s</span></span>`;
-                        }
-                    }
-
-                    if (buffer_diarization) {
-                        if (isFinalizing) {
-                            currentLineText += (currentLineText.length > 0 && buffer_diarization.trim().length > 0 ? " " : "") + buffer_diarization.trim();
-                        } else {
-                            currentLineText += `<span class="buffer_diarization">${buffer_diarization}</span>`;
-                        }
-                    }
-                    if (buffer_transcription) {
-                        if (isFinalizing) {
-                            currentLineText += (currentLineText.length > 0 && buffer_transcription.trim().length > 0 ? " " : "") + buffer_transcription.trim();
-                        } else {
-                            currentLineText += `<span class="buffer_transcription">${buffer_transcription}</span>`;
-                        }
-                    }
-                }
-                
-                return currentLineText.trim().length > 0 || speakerLabel.length > 0
-                    ? `<p>${speakerLabel}<br/><div class='textcontent'>${currentLineText}</div></p>`
-                    : `<p>${speakerLabel}<br/></p>`; 
-            }).join("");
-
-            linesTranscriptDiv.innerHTML = linesHtml;
-        }
-
-        function updateTimer() {
-            if (!startTime) return;
-            
-            const elapsed = Math.floor((Date.now() - startTime) / 1000);
-            const minutes = Math.floor(elapsed / 60).toString().padStart(2, "0");
-            const seconds = (elapsed % 60).toString().padStart(2, "0");
-            timerElement.textContent = `${minutes}:${seconds}`;
-        }
-
-        function drawWaveform() {
-            if (!analyser) return;
-            
-            const bufferLength = analyser.frequencyBinCount;
-            const dataArray = new Uint8Array(bufferLength);
-            analyser.getByteTimeDomainData(dataArray);
-            
-            waveCtx.clearRect(0, 0, waveCanvas.width / (window.devicePixelRatio || 1), waveCanvas.height / (window.devicePixelRatio || 1));
-            waveCtx.lineWidth = 1;
-            waveCtx.strokeStyle = 'rgb(0, 0, 0)';
-            waveCtx.beginPath();
-            
-            const sliceWidth = (waveCanvas.width / (window.devicePixelRatio || 1)) / bufferLength;
-            let x = 0;
-            
-            for (let i = 0; i < bufferLength; i++) {
-                const v = dataArray[i] / 128.0;
-                const y = v * (waveCanvas.height / (window.devicePixelRatio || 1)) / 2;
-                
-                if (i === 0) {
-                    waveCtx.moveTo(x, y);
-                } else {
-                    waveCtx.lineTo(x, y);
-                }
-                
-                x += sliceWidth;
-            }
-            
-            waveCtx.lineTo(waveCanvas.width / (window.devicePixelRatio || 1), waveCanvas.height / (window.devicePixelRatio || 1) / 2);
-            waveCtx.stroke();
-            
-            animationFrame = requestAnimationFrame(drawWaveform);
-        }
-
-        async function startRecording() {
-            try {
-                const stream = await navigator.mediaDevices.getUserMedia({ audio: true });
-                
-                audioContext = new (window.AudioContext || window.webkitAudioContext)();
-                analyser = audioContext.createAnalyser();
-                analyser.fftSize = 256;
-                microphone = audioContext.createMediaStreamSource(stream);
-                microphone.connect(analyser);
-                
-                recorder = new MediaRecorder(stream, { mimeType: "audio/webm" });
-                recorder.ondataavailable = (e) => {
-                    if (websocket && websocket.readyState === WebSocket.OPEN) {
-                        websocket.send(e.data);
-                    }
-                };
-                recorder.start(chunkDuration);
-                
-                startTime = Date.now();
-                timerInterval = setInterval(updateTimer, 1000);
-                drawWaveform();
-                
-                isRecording = true;
-                updateUI();
-            } catch (err) {
-                statusText.textContent = "Error accessing microphone. Please allow microphone access.";
-                console.error(err);
-            }
-        }
-
-        async function stopRecording() {
-            userClosing = true;
-            waitingForStop = true;
-            
-            if (websocket && websocket.readyState === WebSocket.OPEN) {
-                // Send empty audio buffer as stop signal
-                const emptyBlob = new Blob([], { type: 'audio/webm' });
-                websocket.send(emptyBlob);
-                statusText.textContent = "Recording stopped. Processing final audio...";
-            }
-            
-            if (recorder) {
-                recorder.stop();
-                recorder = null;
-            }
-            
-            if (microphone) {
-                microphone.disconnect();
-                microphone = null;
-            }
-            
-            if (analyser) {
-                analyser = null;
-            }
-            
-            if (audioContext && audioContext.state !== 'closed') {
-                try {
-                    audioContext.close();
-                } catch (e) {
-                    console.warn("Could not close audio context:", e);
-                }
-                audioContext = null;
-            }
-            
-            if (animationFrame) {
-                cancelAnimationFrame(animationFrame);
-                animationFrame = null;
-            }
-            
-            if (timerInterval) {
-                clearInterval(timerInterval);
-                timerInterval = null;
-            }            
-            timerElement.textContent = "00:00";
-            startTime = null;
-            
-            
-            isRecording = false;
-            updateUI();	
-        }
-
-        async function toggleRecording() {
-            if (!isRecording) {
-                if (waitingForStop) {
-                    console.log("Waiting for stop, early return");
-                    return;  // Early return, UI is already updated
-                }
-                console.log("Connecting to WebSocket");
-                try {
-                    // If we have an active WebSocket that's still processing, just restart audio capture
-                    if (websocket && websocket.readyState === WebSocket.OPEN) {
-                        await startRecording();
-                    } else {
-                        // If no active WebSocket or it's closed, create new one
-                        await setupWebSocket();
-                        await startRecording();
-                    }
-                } catch (err) {
-                    statusText.textContent = "Could not connect to WebSocket or access mic. Aborted.";
-                    console.error(err);
-                }
-            } else {
-                console.log("Stopping recording");
-                stopRecording();
-            }
-        }
-
-        function updateUI() {
-            recordButton.classList.toggle("recording", isRecording);
-            recordButton.disabled = waitingForStop;
-
-            if (waitingForStop) {
-                if (statusText.textContent !== "Recording stopped. Processing final audio...") {
-                     statusText.textContent = "Please wait for processing to complete...";
-                }
-            } else if (isRecording) {
-                statusText.textContent = "Recording...";
-            } else {
-                if (statusText.textContent !== "Finished processing audio! Ready to record again." &&
-                    statusText.textContent !== "Processing finalized or connection closed.") {
-                    statusText.textContent = "Click to start transcription";
-                }
-            }
-            if (!waitingForStop) {
-                recordButton.disabled = false;
-            }
-        }
-
-        recordButton.addEventListener("click", toggleRecording);
-    </script>
+  <script src="/web/live_transcription.js"></script>
 </body>
-
 </html>

whisperlivekit/web/live_transcription.js

@@ -0,0 +1,513 @@
+/* Theme, WebSocket, recording, rendering logic extracted from inline script and adapted for segmented theme control and WS caption */
+
+let isRecording = false;
+let websocket = null;
+let recorder = null;
+let chunkDuration = 100;
+let websocketUrl = "ws://localhost:8000/asr";
+let userClosing = false;
+let wakeLock = null;
+let startTime = null;
+let timerInterval = null;
+let audioContext = null;
+let analyser = null;
+let microphone = null;
+let waveCanvas = document.getElementById("waveCanvas");
+let waveCtx = waveCanvas.getContext("2d");
+let animationFrame = null;
+let waitingForStop = false;
+let lastReceivedData = null;
+let lastSignature = null;
+
+waveCanvas.width = 60 * (window.devicePixelRatio || 1);
+waveCanvas.height = 30 * (window.devicePixelRatio || 1);
+waveCtx.scale(window.devicePixelRatio || 1, window.devicePixelRatio || 1);
+
+const statusText = document.getElementById("status");
+const recordButton = document.getElementById("recordButton");
+const chunkSelector = document.getElementById("chunkSelector");
+const websocketInput = document.getElementById("websocketInput");
+const websocketDefaultSpan = document.getElementById("wsDefaultUrl");
+const linesTranscriptDiv = document.getElementById("linesTranscript");
+const timerElement = document.querySelector(".timer");
+const themeRadios = document.querySelectorAll('input[name="theme"]');
+
+function getWaveStroke() {
+  const styles = getComputedStyle(document.documentElement);
+  const v = styles.getPropertyValue("--wave-stroke").trim();
+  return v || "#000";
+}
+
+let waveStroke = getWaveStroke();
+function updateWaveStroke() {
+  waveStroke = getWaveStroke();
+}
+
+function applyTheme(pref) {
+  if (pref === "light") {
+    document.documentElement.setAttribute("data-theme", "light");
+  } else if (pref === "dark") {
+    document.documentElement.setAttribute("data-theme", "dark");
+  } else {
+    document.documentElement.removeAttribute("data-theme");
+  }
+  updateWaveStroke();
+}
+
+// Persisted theme preference
+const savedThemePref = localStorage.getItem("themePreference") || "system";
+applyTheme(savedThemePref);
+if (themeRadios.length) {
+  themeRadios.forEach((r) => {
+    r.checked = r.value === savedThemePref;
+    r.addEventListener("change", () => {
+      if (r.checked) {
+        localStorage.setItem("themePreference", r.value);
+        applyTheme(r.value);
+      }
+    });
+  });
+}
+
+// React to OS theme changes when in "system" mode
+const darkMq = window.matchMedia && window.matchMedia("(prefers-color-scheme: dark)");
+const handleOsThemeChange = () => {
+  const pref = localStorage.getItem("themePreference") || "system";
+  if (pref === "system") updateWaveStroke();
+};
+if (darkMq && darkMq.addEventListener) {
+  darkMq.addEventListener("change", handleOsThemeChange);
+} else if (darkMq && darkMq.addListener) {
+  // deprecated, but included for Safari compatibility
+  darkMq.addListener(handleOsThemeChange);
+}
+
+// Helpers
+function fmt1(x) {
+  const n = Number(x);
+  return Number.isFinite(n) ? n.toFixed(1) : x;
+}
+
+// Default WebSocket URL computation
+const host = window.location.hostname || "localhost";
+const port = window.location.port;
+const protocol = window.location.protocol === "https:" ? "wss" : "ws";
+const defaultWebSocketUrl = `${protocol}://${host}${port ? ":" + port : ""}/asr`;
+
+// Populate default caption and input
+if (websocketDefaultSpan) websocketDefaultSpan.textContent = defaultWebSocketUrl;
+websocketInput.value = defaultWebSocketUrl;
+websocketUrl = defaultWebSocketUrl;
+
+// Optional chunk selector (guard for presence)
+if (chunkSelector) {
+  chunkSelector.addEventListener("change", () => {
+    chunkDuration = parseInt(chunkSelector.value);
+  });
+}
+
+// WebSocket input change handling
+websocketInput.addEventListener("change", () => {
+  const urlValue = websocketInput.value.trim();
+  if (!urlValue.startsWith("ws://") && !urlValue.startsWith("wss://")) {
+    statusText.textContent = "Invalid WebSocket URL (must start with ws:// or wss://)";
+    return;
+  }
+  websocketUrl = urlValue;
+  statusText.textContent = "WebSocket URL updated. Ready to connect.";
+});
+
+function setupWebSocket() {
+  return new Promise((resolve, reject) => {
+    try {
+      websocket = new WebSocket(websocketUrl);
+    } catch (error) {
+      statusText.textContent = "Invalid WebSocket URL. Please check and try again.";
+      reject(error);
+      return;
+    }
+
+    websocket.onopen = () => {
+      statusText.textContent = "Connected to server.";
+      resolve();
+    };
+
+    websocket.onclose = () => {
+      if (userClosing) {
+        if (waitingForStop) {
+          statusText.textContent = "Processing finalized or connection closed.";
+          if (lastReceivedData) {
+            renderLinesWithBuffer(
+              lastReceivedData.lines || [],
+              lastReceivedData.buffer_diarization || "",
+              lastReceivedData.buffer_transcription || "",
+              0,
+              0,
+              true
+            );
+          }
+        }
+      } else {
+        statusText.textContent = "Disconnected from the WebSocket server. (Check logs if model is loading.)";
+        if (isRecording) {
+          stopRecording();
+        }
+      }
+      isRecording = false;
+      waitingForStop = false;
+      userClosing = false;
+      lastReceivedData = null;
+      websocket = null;
+      updateUI();
+    };
+
+    websocket.onerror = () => {
+      statusText.textContent = "Error connecting to WebSocket.";
+      reject(new Error("Error connecting to WebSocket"));
+    };
+
+    websocket.onmessage = (event) => {
+      const data = JSON.parse(event.data);
+
+      if (data.type === "ready_to_stop") {
+        console.log("Ready to stop received, finalizing display and closing WebSocket.");
+        waitingForStop = false;
+
+        if (lastReceivedData) {
+          renderLinesWithBuffer(
+            lastReceivedData.lines || [],
+            lastReceivedData.buffer_diarization || "",
+            lastReceivedData.buffer_transcription || "",
+            0,
+            0,
+            true
+          );
+        }
+        statusText.textContent = "Finished processing audio! Ready to record again.";
+        recordButton.disabled = false;
+
+        if (websocket) {
+          websocket.close();
+        }
+        return;
+      }
+
+      lastReceivedData = data;
+
+      const {
+        lines = [],
+        buffer_transcription = "",
+        buffer_diarization = "",
+        remaining_time_transcription = 0,
+        remaining_time_diarization = 0,
+        status = "active_transcription",
+      } = data;
+
+      renderLinesWithBuffer(
+        lines,
+        buffer_diarization,
+        buffer_transcription,
+        remaining_time_diarization,
+        remaining_time_transcription,
+        false,
+        status
+      );
+    };
+  });
+}
+
+function renderLinesWithBuffer(
+  lines,
+  buffer_diarization,
+  buffer_transcription,
+  remaining_time_diarization,
+  remaining_time_transcription,
+  isFinalizing = false,
+  current_status = "active_transcription"
+) {
+  if (current_status === "no_audio_detected") {
+    linesTranscriptDiv.innerHTML =
+      "<p style='text-align: center; color: var(--muted); margin-top: 20px;'><em>No audio detected...</em></p>";
+    return;
+  }
+
+  const showLoading = !isFinalizing && (lines || []).some((it) => it.speaker == 0);
+  const showTransLag = !isFinalizing && remaining_time_transcription > 0;
+  const showDiaLag = !isFinalizing && !!buffer_diarization && remaining_time_diarization > 0;
+  const signature = JSON.stringify({
+    lines: (lines || []).map((it) => ({ speaker: it.speaker, text: it.text, beg: it.beg, end: it.end })),
+    buffer_transcription: buffer_transcription || "",
+    buffer_diarization: buffer_diarization || "",
+    status: current_status,
+    showLoading,
+    showTransLag,
+    showDiaLag,
+    isFinalizing: !!isFinalizing,
+  });
+  if (lastSignature === signature) {
+    const t = document.querySelector(".lag-transcription-value");
+    if (t) t.textContent = fmt1(remaining_time_transcription);
+    const d = document.querySelector(".lag-diarization-value");
+    if (d) d.textContent = fmt1(remaining_time_diarization);
+    const ld = document.querySelector(".loading-diarization-value");
+    if (ld) ld.textContent = fmt1(remaining_time_diarization);
+    return;
+  }
+  lastSignature = signature;
+
+  const linesHtml = (lines || [])
+    .map((item, idx) => {
+      let timeInfo = "";
+      if (item.beg !== undefined && item.end !== undefined) {
+        timeInfo = ` ${item.beg} - ${item.end}`;
+      }
+
+      let speakerLabel = "";
+      if (item.speaker === -2) {
+        speakerLabel = `<span class="silence">Silence<span id='timeInfo'>${timeInfo}</span></span>`;
+      } else if (item.speaker == 0 && !isFinalizing) {
+        speakerLabel = `<span class='loading'><span class="spinner"></span><span id='timeInfo'><span class="loading-diarization-value">${fmt1(
+          remaining_time_diarization
+        )}</span> second(s) of audio are undergoing diarization</span></span>`;
+      } else if (item.speaker !== 0) {
+        speakerLabel = `<span id="speaker">Speaker ${item.speaker}<span id='timeInfo'>${timeInfo}</span></span>`;
+      }
+
+      let currentLineText = item.text || "";
+
+      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">${fmt1(
+              remaining_time_transcription
+            )}</span>s</span></span>`;
+          }
+          if (buffer_diarization && remaining_time_diarization > 0) {
+            speakerLabel += `<span class="label_diarization"><span class="spinner"></span>Diarization lag<span id='timeInfo'><span class="lag-diarization-value">${fmt1(
+              remaining_time_diarization
+            )}</span>s</span></span>`;
+          }
+        }
+
+        if (buffer_diarization) {
+          if (isFinalizing) {
+            currentLineText +=
+              (currentLineText.length > 0 && buffer_diarization.trim().length > 0 ? " " : "") + buffer_diarization.trim();
+          } else {
+            currentLineText += `<span class="buffer_diarization">${buffer_diarization}</span>`;
+          }
+        }
+        if (buffer_transcription) {
+          if (isFinalizing) {
+            currentLineText +=
+              (currentLineText.length > 0 && buffer_transcription.trim().length > 0 ? " " : "") +
+              buffer_transcription.trim();
+          } else {
+            currentLineText += `<span class="buffer_transcription">${buffer_transcription}</span>`;
+          }
+        }
+      }
+
+      return currentLineText.trim().length > 0 || speakerLabel.length > 0
+        ? `<p>${speakerLabel}<br/><div class='textcontent'>${currentLineText}</div></p>`
+        : `<p>${speakerLabel}<br/></p>`;
+    })
+    .join("");
+
+  linesTranscriptDiv.innerHTML = linesHtml;
+  window.scrollTo({ top: document.body.scrollHeight, behavior: "smooth" });
+}
+
+function updateTimer() {
+  if (!startTime) return;
+
+  const elapsed = Math.floor((Date.now() - startTime) / 1000);
+  const minutes = Math.floor(elapsed / 60).toString().padStart(2, "0");
+  const seconds = (elapsed % 60).toString().padStart(2, "0");
+  timerElement.textContent = `${minutes}:${seconds}`;
+}
+
+function drawWaveform() {
+  if (!analyser) return;
+
+  const bufferLength = analyser.frequencyBinCount;
+  const dataArray = new Uint8Array(bufferLength);
+  analyser.getByteTimeDomainData(dataArray);
+
+  waveCtx.clearRect(
+    0,
+    0,
+    waveCanvas.width / (window.devicePixelRatio || 1),
+    waveCanvas.height / (window.devicePixelRatio || 1)
+  );
+  waveCtx.lineWidth = 1;
+  waveCtx.strokeStyle = waveStroke;
+  waveCtx.beginPath();
+
+  const sliceWidth = (waveCanvas.width / (window.devicePixelRatio || 1)) / bufferLength;
+  let x = 0;
+
+  for (let i = 0; i < bufferLength; i++) {
+    const v = dataArray[i] / 128.0;
+    const y = (v * (waveCanvas.height / (window.devicePixelRatio || 1))) / 2;
+
+    if (i === 0) {
+      waveCtx.moveTo(x, y);
+    } else {
+      waveCtx.lineTo(x, y);
+    }
+
+    x += sliceWidth;
+  }
+
+  waveCtx.lineTo(
+    waveCanvas.width / (window.devicePixelRatio || 1),
+    (waveCanvas.height / (window.devicePixelRatio || 1)) / 2
+  );
+  waveCtx.stroke();
+
+  animationFrame = requestAnimationFrame(drawWaveform);
+}
+
+async function startRecording() {
+  try {
+    try {
+      wakeLock = await navigator.wakeLock.request("screen");
+    } catch (err) {
+      console.log("Error acquiring wake lock.");
+    }
+
+    const stream = await navigator.mediaDevices.getUserMedia({ audio: true });
+
+    audioContext = new (window.AudioContext || window.webkitAudioContext)();
+    analyser = audioContext.createAnalyser();
+    analyser.fftSize = 256;
+    microphone = audioContext.createMediaStreamSource(stream);
+    microphone.connect(analyser);
+
+    recorder = new MediaRecorder(stream, { mimeType: "audio/webm" });
+    recorder.ondataavailable = (e) => {
+      if (websocket && websocket.readyState === WebSocket.OPEN) {
+        websocket.send(e.data);
+      }
+    };
+    recorder.start(chunkDuration);
+
+    startTime = Date.now();
+    timerInterval = setInterval(updateTimer, 1000);
+    drawWaveform();
+
+    isRecording = true;
+    updateUI();
+  } catch (err) {
+    statusText.textContent = "Error accessing microphone. Please allow microphone access.";
+    console.error(err);
+  }
+}
+
+async function stopRecording() {
+  if (wakeLock) {
+    try {
+      await wakeLock.release();
+    } catch (e) {
+      // ignore
+    }
+    wakeLock = null;
+  }
+
+  userClosing = true;
+  waitingForStop = true;
+
+  if (websocket && websocket.readyState === WebSocket.OPEN) {
+    const emptyBlob = new Blob([], { type: "audio/webm" });
+    websocket.send(emptyBlob);
+    statusText.textContent = "Recording stopped. Processing final audio...";
+  }
+
+  if (recorder) {
+    recorder.stop();
+    recorder = null;
+  }
+
+  if (microphone) {
+    microphone.disconnect();
+    microphone = null;
+  }
+
+  if (analyser) {
+    analyser = null;
+  }
+
+  if (audioContext && audioContext.state !== "closed") {
+    try {
+      await audioContext.close();
+    } catch (e) {
+      console.warn("Could not close audio context:", e);
+    }
+    audioContext = null;
+  }
+
+  if (animationFrame) {
+    cancelAnimationFrame(animationFrame);
+    animationFrame = null;
+  }
+
+  if (timerInterval) {
+    clearInterval(timerInterval);
+    timerInterval = null;
+  }
+  timerElement.textContent = "00:00";
+  startTime = null;
+
+  isRecording = false;
+  updateUI();
+}
+
+async function toggleRecording() {
+  if (!isRecording) {
+    if (waitingForStop) {
+      console.log("Waiting for stop, early return");
+      return;
+    }
+    console.log("Connecting to WebSocket");
+    try {
+      if (websocket && websocket.readyState === WebSocket.OPEN) {
+        await startRecording();
+      } else {
+        await setupWebSocket();
+        await startRecording();
+      }
+    } catch (err) {
+      statusText.textContent = "Could not connect to WebSocket or access mic. Aborted.";
+      console.error(err);
+    }
+  } else {
+    console.log("Stopping recording");
+    stopRecording();
+  }
+}
+
+function updateUI() {
+  recordButton.classList.toggle("recording", isRecording);
+  recordButton.disabled = waitingForStop;
+
+  if (waitingForStop) {
+    if (statusText.textContent !== "Recording stopped. Processing final audio...") {
+      statusText.textContent = "Please wait for processing to complete...";
+    }
+  } else if (isRecording) {
+    statusText.textContent = "Recording...";
+  } else {
+    if (
+      statusText.textContent !== "Finished processing audio! Ready to record again." &&
+      statusText.textContent !== "Processing finalized or connection closed."
+    ) {
+      statusText.textContent = "Click to start transcription";
+    }
+  }
+  if (!waitingForStop) {
+    recordButton.disabled = false;
+  }
+}
+
+recordButton.addEventListener("click", toggleRecording);

whisperlivekit/web/src/dark_mode.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" height="24px" viewBox="0 -960 960 960" width="24px" fill="#5f6368"><path d="M480-120q-151 0-255.5-104.5T120-480q0-138 90-239.5T440-838q13-2 23 3.5t16 14.5q6 9 6.5 21t-7.5 23q-17 26-25.5 55t-8.5 61q0 90 63 153t153 63q31 0 61.5-9t54.5-25q11-7 22.5-6.5T819-479q10 5 15.5 15t3.5 24q-14 138-117.5 229T480-120Zm0-80q88 0 158-48.5T740-375q-20 5-40 8t-40 3q-123 0-209.5-86.5T364-660q0-20 3-40t8-40q-78 32-126.5 102T200-480q0 116 82 198t198 82Zm-10-270Z"/></svg>

whisperlivekit/web/src/light_mode.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" height="24px" viewBox="0 -960 960 960" width="24px" fill="#5f6368"><path d="M480-360q50 0 85-35t35-85q0-50-35-85t-85-35q-50 0-85 35t-35 85q0 50 35 85t85 35Zm0 80q-83 0-141.5-58.5T280-480q0-83 58.5-141.5T480-680q83 0 141.5 58.5T680-480q0 83-58.5 141.5T480-280ZM80-440q-17 0-28.5-11.5T40-480q0-17 11.5-28.5T80-520h80q17 0 28.5 11.5T200-480q0 17-11.5 28.5T160-440H80Zm720 0q-17 0-28.5-11.5T760-480q0-17 11.5-28.5T800-520h80q17 0 28.5 11.5T920-480q0 17-11.5 28.5T880-440h-80ZM480-760q-17 0-28.5-11.5T440-800v-80q0-17 11.5-28.5T480-920q17 0 28.5 11.5T520-880v80q0 17-11.5 28.5T480-760Zm0 720q-17 0-28.5-11.5T440-80v-80q0-17 11.5-28.5T480-200q17 0 28.5 11.5T520-160v80q0 17-11.5 28.5T480-40ZM226-678l-43-42q-12-11-11.5-28t11.5-29q12-12 29-12t28 12l42 43q11 12 11 28t-11 28q-11 12-27.5 11.5T226-678Zm494 495-42-43q-11-12-11-28.5t11-27.5q11-12 27.5-11.5T734-282l43 42q12 11 11.5 28T777-183q-12 12-29 12t-28-12Zm-42-495q-12-11-11.5-27.5T678-734l42-43q11-12 28-11.5t29 11.5q12 12 12 29t-12 28l-43 42q-12 11-28 11t-28-11ZM183-183q-12-12-12-29t12-28l43-42q12-11 28.5-11t27.5 11q12 11 11.5 27.5T282-226l-42 43q-11 12-28 11.5T183-183Zm297-297Z"/></svg>

whisperlivekit/web/src/system_mode.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" height="24px" viewBox="0 -960 960 960" width="24px" fill="#5f6368"><path d="M396-396q-32-32-58.5-67T289-537q-5 14-6.5 28.5T281-480q0 83 58 141t141 58q14 0 28.5-2t28.5-6q-39-22-74-48.5T396-396Zm85 196q-56 0-107-21t-91-61q-40-40-61-91t-21-107q0-51 17-97.5t50-84.5q13-14 32-9.5t27 24.5q21 55 52.5 104t73.5 91q42 42 91 73.5T648-326q20 8 24.5 27t-9.5 32q-38 33-84.5 50T481-200Zm223-192q-16-5-23-20.5t-4-32.5q9-48-6-94.5T621-621q-35-35-80.5-49.5T448-677q-17 3-32-4t-21-23q-6-16 1.5-31t23.5-19q69-15 138 4.5T679-678q51 51 71 120t5 138q-4 17-19 25t-32 3ZM480-840q-17 0-28.5-11.5T440-880v-40q0-17 11.5-28.5T480-960q17 0 28.5 11.5T520-920v40q0 17-11.5 28.5T480-840Zm0 840q-17 0-28.5-11.5T440-40v-40q0-17 11.5-28.5T480-120q17 0 28.5 11.5T520-80v40q0 17-11.5 28.5T480 0Zm255-734q-12-12-12-28.5t12-28.5l28-28q11-11 27.5-11t28.5 11q12 12 12 28.5T819-762l-28 28q-12 12-28 12t-28-12ZM141-141q-12-12-12-28.5t12-28.5l28-28q12-12 28-12t28 12q12 12 12 28.5T225-169l-28 28q-11 11-27.5 11T141-141Zm739-299q-17 0-28.5-11.5T840-480q0-17 11.5-28.5T880-520h40q17 0 28.5 11.5T960-480q0 17-11.5 28.5T920-440h-40Zm-840 0q-17 0-28.5-11.5T0-480q0-17 11.5-28.5T40-520h40q17 0 28.5 11.5T120-480q0 17-11.5 28.5T80-440H40Zm779 299q-12 12-28.5 12T762-141l-28-28q-12-12-12-28t12-28q12-12 28.5-12t28.5 12l28 28q11 11 11 27.5T819-141ZM226-735q-12 12-28.5 12T169-735l-28-28q-11-11-11-27.5t11-28.5q12-12 28.5-12t28.5 12l28 28q12 12 12 28t-12 28Zm170 339Z"/></svg>

whisperlivekit/web/web_interface.py

@@ -10,4 +10,24 @@ def get_web_interface_html():
             return f.read()
     except Exception as e:
         logger.error(f"Error loading web interface HTML: {e}")
-        return "<html><body><h1>Error loading interface</h1></body></html>"
+        return "<html><body><h1>Error loading interface</h1></body></html>"
+
+
+if __name__ == '__main__':
+    
+    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()    
+    web_dir = pathlib.Path(webpkg.__file__).parent
+    app.mount("/web", StaticFiles(directory=str(web_dir)), name="web")
+    
+    @app.get("/")
+    async def get():
+        return HTMLResponse(get_web_interface_html())
+
+    uvicorn.run(app=app)

whisperlivekit/whisper_streaming_custom/backends.py

@@ -3,29 +3,10 @@ import logging
 import io
 import soundfile as sf
 import math
-try: 
-    import torch
-except ImportError: 
-    torch = None
 from typing import List
 import numpy as np
 from whisperlivekit.timed_objects import ASRToken
-
 logger = logging.getLogger(__name__)
-
-try:
-    from whisperlivekit.simul_whisper.config import AlignAttConfig
-    from whisperlivekit.simul_whisper.simul_whisper import PaddedAlignAttWhisper, DEC_PAD
-    from whisperlivekit.simul_whisper.whisper import tokenizer
-    SIMULSTREAMING_AVAILABLE = True
-except ImportError:
-    logger.warning("SimulStreaming dependencies not available. SimulStreaming backend will not be available.")
-    SIMULSTREAMING_AVAILABLE = False
-    AlignAttConfig = None
-    PaddedAlignAttWhisper = None
-    DEC_PAD = None
-    tokenizer = None
-
 class ASRBase:
     sep = " "  # join transcribe words with this character (" " for whisper_timestamped,
               # "" for faster-whisper because it emits the spaces when needed)
@@ -306,181 +287,4 @@ class OpenaiApiASR(ASRBase):
         self.use_vad_opt = True
 
     def set_translate_task(self):
-        self.task = "translate"
-
-
-class SimulStreamingASR(ASRBase):
-    """SimulStreaming backend with AlignAtt policy."""
-    sep = " "
-
-    def __init__(self, lan, modelsize=None, cache_dir=None, model_dir=None, logfile=sys.stderr, **kwargs):
-        if not SIMULSTREAMING_AVAILABLE:
-            raise ImportError("""SimulStreaming dependencies are not available. Please install WhisperLiveKit using pip install "whisperlivekit[simulstreaming]". If you are building from source, you should also copy the content of the simul_whisper directory from the SimulStreaming repository into whisperlivekit/simul_whisper.""")
-        with open("whisperlivekit/simul_whisper/dual_license_simulstreaming.md", "r") as f:
-            print("*"*80 + f.read() + "*"*80)
-        self.logfile = logfile
-        self.transcribe_kargs = {}
-        self.original_language = None if lan == "auto" else lan
-        
-        self.model_path = kwargs.get('model_path', './large-v3.pt')
-        self.frame_threshold = kwargs.get('frame_threshold', 25)
-        self.audio_max_len = kwargs.get('audio_max_len', 30.0)
-        self.audio_min_len = kwargs.get('audio_min_len', 0.0)
-        self.segment_length = kwargs.get('segment_length', 0.5)
-        self.beams = kwargs.get('beams', 1)
-        self.decoder_type = kwargs.get('decoder_type', 'greedy' if self.beams == 1 else 'beam')
-        self.task = kwargs.get('task', 'transcribe')
-        self.cif_ckpt_path = kwargs.get('cif_ckpt_path', None)
-        self.never_fire = kwargs.get('never_fire', False)
-        self.init_prompt = kwargs.get('init_prompt', None)
-        self.static_init_prompt = kwargs.get('static_init_prompt', None)
-        self.max_context_tokens = kwargs.get('max_context_tokens', None)
-        
-        if model_dir is not None:
-            self.model_path = model_dir
-        elif modelsize is not None: #For the moment the .en.pt models do not work!
-            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_path = model_mapping.get(modelsize, f'./{modelsize}.pt')
-        
-        self.model = self.load_model(modelsize, cache_dir, model_dir)
-        
-        # Set up tokenizer for translation if needed
-        if self.task == "translate":
-            self.set_translate_task()
-
-    def load_model(self, modelsize, cache_dir, model_dir):
-        try:
-            cfg = AlignAttConfig(
-                model_path=self.model_path,
-                segment_length=self.segment_length,
-                frame_threshold=self.frame_threshold,
-                language=self.original_language,
-                audio_max_len=self.audio_max_len,
-                audio_min_len=self.audio_min_len,
-                cif_ckpt_path=self.cif_ckpt_path,
-                decoder_type="beam",
-                beam_size=self.beams,
-                task=self.task,
-                never_fire=self.never_fire,
-                init_prompt=self.init_prompt,
-                max_context_tokens=self.max_context_tokens,
-                static_init_prompt=self.static_init_prompt,
-            )
-            
-            logger.info(f"Loading SimulStreaming model with language: {self.original_language}")
-            model = PaddedAlignAttWhisper(cfg)
-            return model
-            
-        except Exception as e:
-            logger.error(f"Failed to load SimulStreaming model: {e}")
-            raise
-
-    def transcribe(self, audio, init_prompt=""):
-        """Transcribe audio using SimulStreaming."""
-        try:
-            if isinstance(audio, np.ndarray):
-                audio_tensor = torch.from_numpy(audio).float()
-            else:
-                audio_tensor = audio
-                
-            prompt = init_prompt if init_prompt else (self.init_prompt or "")
-            
-            result = self.model.infer(audio_tensor, init_prompt=prompt)
-            
-            if torch.is_tensor(result):
-                result = result[result < DEC_PAD]
-                
-            logger.debug(f"SimulStreaming transcription result: {result}")
-            return result
-            
-        except Exception as e:
-            logger.error(f"SimulStreaming transcription failed: {e}")
-            raise
-
-    def ts_words(self, result) -> List[ASRToken]:
-        """Convert SimulStreaming result to ASRToken list."""
-        tokens = []
-        
-        try:
-            if torch.is_tensor(result):
-                text = self.model.tokenizer.decode(result.cpu().numpy())
-            else:
-                text = str(result)
-                
-            if not text or len(text.strip()) == 0:
-                return tokens
-                
-            # We dont have word-level timestamps here. 1rst approach, should be improved later.
-            words = text.strip().split()
-            if not words:
-                return tokens
-            
-            duration_per_word = 0.1  # this will be modified based on actual audio duration 
-            #with the SimulStreamingOnlineProcessor
-            
-            for i, word in enumerate(words):
-                start_time = i * duration_per_word
-                end_time = (i + 1) * duration_per_word
-                
-                token = ASRToken(
-                    start=start_time,
-                    end=end_time,
-                    text=word,
-                    probability=1.0
-                )
-                tokens.append(token)
-                
-        except Exception as e:
-            logger.error(f"Error converting SimulStreaming result to tokens: {e}")
-            
-        return tokens
-
-    def segments_end_ts(self, result) -> List[float]:
-        """Get segment end timestamps."""
-        if torch.is_tensor(result):
-            num_tokens = len(result)
-            return [num_tokens * 0.1]  # rough estimate
-        return [1.0]
-
-    def use_vad(self):
-        """Enable VAD - SimulStreaming has different VAD handling."""
-        logger.info("VAD requested for SimulStreaming - handled internally by the model")
-        pass
-
-    def set_translate_task(self):
-        """Set up translation task."""
-        try:
-            self.model.tokenizer = tokenizer.get_tokenizer(
-                multilingual=True,
-                language=self.model.cfg.language,
-                num_languages=self.model.model.num_languages,
-                task="translate"
-            )
-            logger.info("SimulStreaming configured for translation task")
-        except Exception as e:
-            logger.error(f"Failed to configure SimulStreaming for translation: {e}")
-            raise
-
-    def warmup(self, audio, init_prompt=""):
-        """Warmup the SimulStreaming model."""
-        try:
-            if isinstance(audio, np.ndarray):
-                audio = torch.from_numpy(audio).float()
-            self.model.infer(audio, True)
-            self.model.refresh_segment(complete=True)
-            logger.info("SimulStreaming model warmed up successfully")
-        except Exception as e:
-            logger.warning(f"SimulStreaming warmup failed: {e}")
+        self.task = "translate"

whisperlivekit/whisper_streaming_custom/online_asr.py

@@ -6,18 +6,6 @@ from whisperlivekit.timed_objects import ASRToken, Sentence, Transcript
 
 logger = logging.getLogger(__name__)
 
-# simulStreaming imports - we check if the files are here
-try:
-    import torch
-    from simul_whisper.config import AlignAttConfig
-    SIMULSTREAMING_AVAILABLE = True
-except ImportError:
-    logger.warning("SimulStreaming dependencies not available for online processor.")
-    SIMULSTREAMING_AVAILABLE = False
-    OnlineProcessorInterface = None
-    torch = None
-
-
 class HypothesisBuffer:
     """
     Buffer to store and process ASR hypothesis tokens.
@@ -134,6 +122,7 @@ class OnlineASRProcessor:
         self.tokenize = tokenize_method
         self.logfile = logfile
         self.confidence_validation = confidence_validation
+        self.global_time_offset = 0.0
         self.init()
 
         self.buffer_trimming_way, self.buffer_trimming_sec = buffer_trimming
@@ -154,6 +143,7 @@ class OnlineASRProcessor:
         self.buffer_time_offset = offset if offset is not None else 0.0
         self.transcript_buffer.last_committed_time = self.buffer_time_offset
         self.committed: List[ASRToken] = []
+        self.time_of_last_asr_output = 0.0
 
     def get_audio_buffer_end_time(self) -> float:
         """Returns the absolute end time of the current audio_buffer."""
@@ -163,6 +153,21 @@ 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)
+        else:
+            self.init(offset=silence_duration + offset)
+        self.global_time_offset += silence_duration
+
     def prompt(self) -> Tuple[str, str]:
         """
         Returns a tuple: (prompt, context), where:
@@ -210,11 +215,26 @@ class OnlineASRProcessor:
         self.transcript_buffer.insert(tokens, self.buffer_time_offset)
         committed_tokens = self.transcript_buffer.flush()
         self.committed.extend(committed_tokens)
+
+        if committed_tokens:
+            self.time_of_last_asr_output = self.committed[-1].end
+
         completed = self.concatenate_tokens(committed_tokens)
         logger.debug(f">>>> COMPLETE NOW: {completed.text}")
         incomp = self.concatenate_tokens(self.transcript_buffer.buffer)
         logger.debug(f"INCOMPLETE: {incomp.text}")
 
+        buffer_duration = len(self.audio_buffer) / self.SAMPLING_RATE
+        if not committed_tokens and buffer_duration > self.buffer_trimming_sec:
+            time_since_last_output = self.get_audio_buffer_end_time() - self.time_of_last_asr_output
+            if time_since_last_output > self.buffer_trimming_sec:
+                logger.warning(
+                    f"No ASR output for {time_since_last_output:.2f}s. "
+                    f"Resetting buffer to prevent freezing."
+                )
+                self.init(offset=self.get_audio_buffer_end_time())
+                return [], current_audio_processed_upto
+
         if committed_tokens and self.buffer_trimming_way == "sentence":
             if len(self.audio_buffer) / self.SAMPLING_RATE > self.buffer_trimming_sec:
                 self.chunk_completed_sentence()
@@ -226,6 +246,9 @@ class OnlineASRProcessor:
         logger.debug(
             f"Length of audio buffer now: {len(self.audio_buffer)/self.SAMPLING_RATE:.2f} seconds"
         )
+        if self.global_time_offset:
+            for token in committed_tokens:
+                token = token.with_offset(self.global_time_offset)
         return committed_tokens, current_audio_processed_upto
 
     def chunk_completed_sentence(self):
@@ -387,331 +410,3 @@ class OnlineASRProcessor:
             start = None
             end = None
         return Transcript(start, end, text, probability=probability)
-
-
-class VACOnlineASRProcessor:
-    """
-    Wraps an OnlineASRProcessor with a Voice Activity Controller (VAC).
-    
-    It receives small chunks of audio, applies VAD (e.g. with Silero),
-    and when the system detects a pause in speech (or end of an utterance)
-    it finalizes the utterance immediately.
-    """
-    SAMPLING_RATE = 16000
-
-    def __init__(self, online_chunk_size: float, *args, **kwargs):
-        self.online_chunk_size = online_chunk_size
-        self.online = OnlineASRProcessor(*args, **kwargs)
-        self.asr = self.online.asr
-        
-        # Load a VAD model (e.g. Silero VAD)
-        import torch
-        model, _ = torch.hub.load(repo_or_dir="snakers4/silero-vad", model="silero_vad")
-        from .silero_vad_iterator import FixedVADIterator
-
-        self.vac = FixedVADIterator(model)
-        self.logfile = self.online.logfile
-        self.last_input_audio_stream_end_time: float = 0.0
-        self.init()
-
-    def init(self):
-        self.online.init()
-        self.vac.reset_states()
-        self.current_online_chunk_buffer_size = 0
-        self.last_input_audio_stream_end_time = self.online.buffer_time_offset
-        self.is_currently_final = False
-        self.status: Optional[str] = None  # "voice" or "nonvoice"
-        self.audio_buffer = np.array([], dtype=np.float32)
-        self.buffer_offset = 0  # in frames
-
-    def get_audio_buffer_end_time(self) -> float:
-        """Returns the absolute end time of the audio processed by the underlying OnlineASRProcessor."""
-        return self.online.get_audio_buffer_end_time()
-
-    def clear_buffer(self):
-        self.buffer_offset += len(self.audio_buffer)
-        self.audio_buffer = np.array([], dtype=np.float32)
-
-    def insert_audio_chunk(self, audio: np.ndarray, audio_stream_end_time: float):
-        """
-        Process an incoming small audio chunk:
-          - run VAD on the chunk,
-          - decide whether to send the audio to the online ASR processor immediately,
-          - and/or to mark the current utterance as finished.
-        """
-        self.last_input_audio_stream_end_time = audio_stream_end_time
-        res = self.vac(audio)
-        self.audio_buffer = np.append(self.audio_buffer, audio)
-
-        if res is not None:
-            # VAD returned a result; adjust the frame number
-            frame = list(res.values())[0] - self.buffer_offset
-            if "start" in res and "end" not in res:
-                self.status = "voice"
-                send_audio = self.audio_buffer[frame:]
-                self.online.init(offset=(frame + self.buffer_offset) / self.SAMPLING_RATE)
-                self.online.insert_audio_chunk(send_audio)
-                self.current_online_chunk_buffer_size += len(send_audio)
-                self.clear_buffer()
-            elif "end" in res and "start" not in res:
-                self.status = "nonvoice"
-                send_audio = self.audio_buffer[:frame]
-                self.online.insert_audio_chunk(send_audio)
-                self.current_online_chunk_buffer_size += len(send_audio)
-                self.is_currently_final = True
-                self.clear_buffer()
-            else:
-                beg = res["start"] - self.buffer_offset
-                end = res["end"] - self.buffer_offset
-                self.status = "nonvoice"
-                send_audio = self.audio_buffer[beg:end]
-                self.online.init(offset=(beg + self.buffer_offset) / self.SAMPLING_RATE)
-                self.online.insert_audio_chunk(send_audio)
-                self.current_online_chunk_buffer_size += len(send_audio)
-                self.is_currently_final = True
-                self.clear_buffer()
-        else:
-            if self.status == "voice":
-                self.online.insert_audio_chunk(self.audio_buffer)
-                self.current_online_chunk_buffer_size += len(self.audio_buffer)
-                self.clear_buffer()
-            else:
-                # Keep 1 second worth of audio in case VAD later detects voice,
-                # but trim to avoid unbounded memory usage.
-                self.buffer_offset += max(0, len(self.audio_buffer) - self.SAMPLING_RATE)
-                self.audio_buffer = self.audio_buffer[-self.SAMPLING_RATE:]
-
-    def process_iter(self) -> Tuple[List[ASRToken], float]:
-        """
-        Depending on the VAD status and the amount of accumulated audio,
-        process the current audio chunk.
-        Returns a tuple: (list of committed ASRToken objects, float representing the audio processed up to time).
-        """
-        if self.is_currently_final:
-            return self.finish()
-        elif self.current_online_chunk_buffer_size > self.SAMPLING_RATE * self.online_chunk_size:
-            self.current_online_chunk_buffer_size = 0
-            return self.online.process_iter()
-        else:
-            logger.debug("No online update, only VAD")
-            return [], self.last_input_audio_stream_end_time
-
-    def finish(self) -> Tuple[List[ASRToken], float]:
-        """
-        Finish processing by flushing any remaining text.
-        Returns a tuple: (list of remaining ASRToken objects, float representing the final audio processed up to time).
-        """
-        result_tokens, processed_upto = self.online.finish()
-        self.current_online_chunk_buffer_size = 0
-        self.is_currently_final = False
-        return result_tokens, processed_upto
-    
-    def get_buffer(self):
-        """
-        Get the unvalidated buffer in string format.
-        """
-        return self.online.concatenate_tokens(self.online.transcript_buffer.buffer)
-
-
-class SimulStreamingOnlineProcessor:
-    SAMPLING_RATE = 16000
-
-    def __init__(
-        self,
-        asr,
-        tokenize_method: Optional[callable] = None,
-        buffer_trimming: Tuple[str, float] = ("segment", 15),
-        confidence_validation = False,
-        logfile=sys.stderr,
-    ):
-        if not SIMULSTREAMING_AVAILABLE:
-            raise ImportError("SimulStreaming dependencies are not available.")
-        
-        self.asr = asr
-        self.tokenize = tokenize_method
-        self.logfile = logfile
-        self.confidence_validation = confidence_validation
-        self.init()
-        
-        # buffer does not work yet
-        self.buffer_trimming_way, self.buffer_trimming_sec = buffer_trimming
-
-    def init(self, offset: Optional[float] = None):
-        """Initialize or reset the processing state."""
-        self.audio_chunks = []
-        self.offset = offset if offset is not None else 0.0
-        self.is_last = False
-        self.beg = self.offset
-        self.end = self.offset
-        self.cumulative_audio_duration = 0.0
-        self.last_audio_stream_end_time = self.offset
-        
-        self.committed: List[ASRToken] = []
-        self.last_result_tokens: List[ASRToken] = []        
-        self.buffer_content = ""
-        self.processed_audio_duration = 0.0
-
-    def get_audio_buffer_end_time(self) -> float:
-        """Returns the absolute end time of the current audio buffer."""
-        return self.end
-
-    def insert_audio_chunk(self, audio: np.ndarray, audio_stream_end_time: Optional[float] = None):
-        """Append an audio chunk to be processed by SimulStreaming."""
-        if torch is None:
-            raise ImportError("PyTorch is required for SimulStreaming but not available")
-            
-        # Convert numpy array to torch tensor
-        audio_tensor = torch.from_numpy(audio).float()
-        self.audio_chunks.append(audio_tensor)
-        
-        # Update timing
-        chunk_duration = len(audio) / self.SAMPLING_RATE
-        self.cumulative_audio_duration += chunk_duration
-        
-        if audio_stream_end_time is not None:
-            self.last_audio_stream_end_time = audio_stream_end_time
-            self.end = audio_stream_end_time
-        else:
-            self.end = self.offset + self.cumulative_audio_duration
-
-    def prompt(self) -> Tuple[str, str]:
-        """
-        Returns a tuple: (prompt, context).
-        SimulStreaming handles prompting internally, so we return empty strings.
-        """
-        return "", ""
-
-    def get_buffer(self):
-        """
-        Get the unvalidated buffer content.
-        """
-        buffer_end = self.end if hasattr(self, 'end') else None
-        return Transcript(
-            start=None, 
-            end=buffer_end, 
-            text=self.buffer_content, 
-            probability=None
-        )
-
-    def process_iter(self) -> Tuple[List[ASRToken], float]:
-        """
-        Process accumulated audio chunks using SimulStreaming.
-        
-        Returns a tuple: (list of committed ASRToken objects, float representing the audio processed up to time).
-        """
-        if not self.audio_chunks:
-            return [], self.end
-
-        try:
-            # concatenate all audio chunks
-            if len(self.audio_chunks) == 1:
-                audio = self.audio_chunks[0]
-            else:
-                audio = torch.cat(self.audio_chunks, dim=0)
-            
-            audio_duration = audio.shape[0] / self.SAMPLING_RATE if audio.shape[0] > 0 else 0
-            self.processed_audio_duration += audio_duration
-            
-            self.audio_chunks = []
-            
-            logger.debug(f"SimulStreaming processing audio shape: {audio.shape}, duration: {audio_duration:.2f}s")
-            logger.debug(f"Current end time: {self.end:.2f}s, last stream time: {self.last_audio_stream_end_time:.2f}s")
-            
-            result = self.asr.model.infer(audio, is_last=self.is_last)
-            
-            if torch.is_tensor(result):
-                # we filter out padding tokens as it s done in simul whisper
-                from simul_whisper.simul_whisper import DEC_PAD
-                result = result[result < DEC_PAD]
-                
-                # C/P from simul_whisper.simul_whisper.py
-                if len(result) > 0:
-                    decoded_text = self.asr.model.tokenizer.decode(result.cpu().numpy())
-                    logger.debug(f"SimulStreaming decoded: {decoded_text}")
-                    
-                    if decoded_text.strip():
-                        words = decoded_text.strip().split()
-                        new_tokens = []
-                        
-                        num_words = len(words)
-                        if num_words > 0:
-                            # distribute words evenly across the processed audio duration
-                            # we NEED that for when we use diarization. Even if that s not perfect
-                            start_time = self.end - audio_duration
-                            time_per_word = audio_duration / num_words if num_words > 1 else audio_duration
-                            
-                            for i, word in enumerate(words):
-                                token_start = start_time + (i * time_per_word)
-                                token_end = start_time + ((i + 1) * time_per_word)
-                                
-                                token_end = min(token_end, self.end)
-                                
-                                token = ASRToken(
-                                    start=token_start,
-                                    end=token_end,
-                                    text=word,
-                                    probability=0.95  # fake prob. Maybe we can extract it from the model?
-                                )
-                                new_tokens.append(token)
-                        
-                        self.beg = self.end
-                        
-                        self.committed.extend(new_tokens)
-                        self.last_result_tokens = new_tokens
-                        
-                        logger.debug(f"SimulStreaming generated {len(new_tokens)} tokens with end time: {self.end:.2f}s")
-                        return new_tokens, self.end
-            
-            return [], self.end
-            
-        except Exception as e:
-            logger.error(f"SimulStreaming processing error: {e}")
-            logger.error(f"Error details: {type(e).__name__}: {str(e)}")
-            return [], self.end
-
-    def finish(self) -> Tuple[List[ASRToken], float]:
-        logger.debug("SimulStreaming finish() called")
-        self.is_last = True        
-        final_tokens, final_time = self.process_iter()
-        self.is_last = False
-        return final_tokens, final_time
-
-    def concatenate_tokens(
-        self,
-        tokens: List[ASRToken],
-        sep: Optional[str] = None,
-        offset: float = 0
-    ) -> Transcript:
-        """Concatenate tokens into a Transcript object."""
-        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
-        if tokens:
-            start = offset + tokens[0].start
-            end = offset + tokens[-1].end
-        else:
-            start = None
-            end = None
-        return Transcript(start, end, text, probability=probability)
-
-    def chunk_at(self, time: float):
-        """
-        useless but kept for compatibility
-        """
-        logger.debug(f"SimulStreaming chunk_at({time:.2f}) - handled internally")
-        pass
-
-    def words_to_sentences(self, tokens: List[ASRToken]) -> List[Sentence]:
-        """
-        Create simple sentences.
-        """
-        if not tokens:
-            return []
-        
-        full_text = " ".join(token.text for token in tokens)
-        sentence = Sentence(
-            start=tokens[0].start,
-            end=tokens[-1].end,
-            text=full_text
-        )
-        return [sentence]

whisperlivekit/whisper_streaming_custom/whisper_online.py

@@ -5,8 +5,7 @@ import librosa
 from functools import lru_cache
 import time
 import logging
-from .backends import FasterWhisperASR, MLXWhisper, WhisperTimestampedASR, OpenaiApiASR, SimulStreamingASR, SIMULSTREAMING_AVAILABLE
-from .online_asr import OnlineASRProcessor, VACOnlineASRProcessor, SimulStreamingOnlineProcessor, SIMULSTREAMING_AVAILABLE as SIMULSTREAMING_ONLINE_AVAILABLE
+from .backends import FasterWhisperASR, MLXWhisper, WhisperTimestampedASR, OpenaiApiASR
 
 logger = logging.getLogger(__name__)
 
@@ -68,38 +67,7 @@ def backend_factory(args):
     backend = args.backend
     if backend == "openai-api":
         logger.debug("Using OpenAI API.")
-        asr = OpenaiApiASR(lan=args.lan)
-    elif backend == "simulstreaming":
-        logger.debug("Using SimulStreaming backend.")
-        if not SIMULSTREAMING_AVAILABLE:
-            raise ImportError(
-                "SimulStreaming backend is not available. Please install SimulStreaming dependencies. "
-                "See the documentation for installation instructions."
-            )
-        
-        simulstreaming_kwargs = {}
-        for attr in ['frame_threshold', 'beams', 'decoder_type', 'audio_max_len', 'audio_min_len', 
-                     'cif_ckpt_path', 'never_fire', 'init_prompt', 'static_init_prompt', 
-                     'max_context_tokens', 'model_path']:
-            if hasattr(args, attr):
-                simulstreaming_kwargs[attr] = getattr(args, attr)
-        
-        # Add segment_length from min_chunk_size
-        simulstreaming_kwargs['segment_length'] = getattr(args, 'min_chunk_size', 0.5)
-        simulstreaming_kwargs['task'] = args.task
-        
-        size = args.model
-        t = time.time()
-        logger.info(f"Loading SimulStreaming {size} model for language {args.lan}...")
-        asr = SimulStreamingASR(
-            modelsize=size,
-            lan=args.lan,
-            cache_dir=getattr(args, 'model_cache_dir', None),
-            model_dir=getattr(args, 'model_dir', None),
-            **simulstreaming_kwargs
-        )
-        e = time.time()
-        logger.info(f"done. It took {round(e-t,2)} seconds.")
+        asr = OpenaiApiASR(lan=args.lan)        
     else:
         if backend == "faster-whisper":
             asr_cls = FasterWhisperASR
@@ -139,107 +107,4 @@ def backend_factory(args):
         tokenizer = create_tokenizer(tgt_language)
     else:
         tokenizer = None
-    return asr, tokenizer
-
-def online_factory(args, asr, tokenizer, logfile=sys.stderr):
-    if args.backend == "simulstreaming":
-        if not SIMULSTREAMING_ONLINE_AVAILABLE:
-            raise ImportError("SimulStreaming online processor is not available.")
-        
-        logger.debug("Creating SimulStreaming online processor")
-        online = SimulStreamingOnlineProcessor(
-            asr,
-            tokenizer,
-            logfile=logfile,
-            buffer_trimming=(args.buffer_trimming, args.buffer_trimming_sec),
-            confidence_validation=args.confidence_validation
-        )
-    elif args.vac:
-        online = VACOnlineASRProcessor(
-            args.min_chunk_size,
-            asr,
-            tokenizer,
-            logfile=logfile,
-            buffer_trimming=(args.buffer_trimming, args.buffer_trimming_sec),
-            confidence_validation = args.confidence_validation
-        )
-    else:
-        online = OnlineASRProcessor(
-            asr,
-            tokenizer,
-            logfile=logfile,
-            buffer_trimming=(args.buffer_trimming, args.buffer_trimming_sec),
-            confidence_validation = args.confidence_validation
-        )
-    return online
-  
-def asr_factory(args, logfile=sys.stderr):
-    """
-    Creates and configures an ASR and ASR Online instance based on the specified backend and arguments.
-    """
-    asr, tokenizer = backend_factory(args)
-    online = online_factory(args, asr, tokenizer, logfile=logfile)
-    return asr, online
-
-def warmup_asr(asr, warmup_file=None, timeout=5):
-    """
-    Warmup the ASR model by transcribing a short audio file.
-    """
-    import os
-    import tempfile
-    
-    is_simulstreaming = hasattr(asr, 'warmup') and callable(getattr(asr, 'warmup'))
-    
-    if warmup_file is None:
-        # Download JFK sample if not already present
-        jfk_url = "https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav"
-        temp_dir = tempfile.gettempdir()
-        warmup_file = os.path.join(temp_dir, "whisper_warmup_jfk.wav")
-        
-        if not os.path.exists(warmup_file):
-            logger.debug(f"Downloading warmup file from {jfk_url}")
-            print(f"Downloading warmup file from {jfk_url}")
-            import time
-            import urllib.request
-            import urllib.error
-            import socket
-            
-            original_timeout = socket.getdefaulttimeout()
-            socket.setdefaulttimeout(timeout)
-            
-            start_time = time.time()
-            try:
-                urllib.request.urlretrieve(jfk_url, warmup_file)
-                logger.debug(f"Download successful in {time.time() - start_time:.2f}s")
-            except (urllib.error.URLError, socket.timeout) as e:
-                logger.warning(f"Download failed: {e}. Proceeding without warmup.")
-                return False
-            finally:
-                socket.setdefaulttimeout(original_timeout)
-    elif not warmup_file:
-        return False 
-    
-    if not warmup_file or not os.path.exists(warmup_file) or os.path.getsize(warmup_file) == 0:
-        logger.warning(f"Warmup file {warmup_file} invalid or missing.")
-        return False
-    
-    print(f"Warming up {'SimulStreaming' if is_simulstreaming else 'Whisper'} with {warmup_file}")
-    try:
-        import librosa
-        audio, sr = librosa.load(warmup_file, sr=16000)
-    except Exception as e:
-        logger.warning(f"Failed to load audio file: {e}")
-        return False
-    
-    try:
-        if is_simulstreaming:
-            asr.warmup(audio)
-        else:
-            asr.transcribe(audio)
-        
-        logger.info(f"{'SimulStreaming' if is_simulstreaming else 'Whisper'} is warmed up")
-        return True
-        
-    except Exception as e:
-        logger.warning(f"Warmup failed: {e}")
-        return False
+    return asr, tokenizer