├── src └── whisperx_api_server │ ├── __init__.py │ ├── routers │ ├── __init__.py │ ├── misc.py │ ├── models.py │ └── transcriptions.py │ ├── .gitignore │ ├── api-keys │ └── keys.json │ ├── logger.py │ ├── dependencies.py │ ├── formatters.py │ ├── main.py │ ├── config.py │ ├── transcriber.py │ └── models.py ├── constraints.txt ├── requirements-cpu.txt ├── requirements-cuda.txt ├── .dockerignore ├── requirements.txt ├── cuda-docker-entrypoint.sh ├── Dockerfile.cpu ├── compose.yaml ├── .github └── workflows │ ├── docker-build-and-push-cpu.yml │ └── docker-build-and-push-cuda.yml ├── Dockerfile.cuda ├── README.md ├── .gitignore └── LICENSE /src/whisperx_api_server/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /src/whisperx_api_server/routers/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /constraints.txt: -------------------------------------------------------------------------------- 1 | torch~=2.6.0 2 | torchaudio~=2.6.0 3 | onnxruntime~=1.21.0 -------------------------------------------------------------------------------- /src/whisperx_api_server/.gitignore: -------------------------------------------------------------------------------- 1 | **/__pycache__/ 2 | venv/ 3 | env/ 4 | .venv/ -------------------------------------------------------------------------------- /src/whisperx_api_server/api-keys/keys.json: -------------------------------------------------------------------------------- 1 | // { 2 | // "some-key": "some-client" 3 | // } -------------------------------------------------------------------------------- /requirements-cpu.txt: -------------------------------------------------------------------------------- 1 | --extra-index-url https://download.pytorch.org/whl/cpu 2 | torch~=2.6.0 3 | torchaudio~=2.6.0 -------------------------------------------------------------------------------- /requirements-cuda.txt: -------------------------------------------------------------------------------- 1 | --extra-index-url https://download.pytorch.org/whl/cu124 2 | torch~=2.6.0 3 | torchaudio~=2.6.0 -------------------------------------------------------------------------------- /.dockerignore: -------------------------------------------------------------------------------- 1 | __pycache__ 2 | .git 3 | .gitignore 4 | README.md 5 | models 6 | venv 7 | scripts 8 | .vscode 9 | whisperx/hf-home 10 | whisperx/torch-home 11 | whisperx/test-files 12 | Dockerfile* 13 | .scannerwork -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | fastapi>=0.120.4 2 | uvicorn>=0.38.0 3 | pydantic>=2.12.3 4 | pydantic-settings>=2.11.0 5 | whisperx @ git+https://github.com/m-bain/whisperX.git@429658d4ccefa55244bcdccd5d179795436093e4 6 | python-multipart>=0.0.20 -------------------------------------------------------------------------------- /src/whisperx_api_server/routers/misc.py: -------------------------------------------------------------------------------- 1 | import logging 2 | from fastapi import APIRouter 3 | from fastapi.responses import JSONResponse 4 | 5 | from whisperx_api_server.config import ( 6 | MediaType, 7 | ) 8 | 9 | logger = logging.getLogger(__name__) 10 | 11 | router = APIRouter() 12 | 13 | @router.get( 14 | "/healthcheck", 15 | description="Check the health of the API server", 16 | tags=["Misc"], 17 | ) 18 | def health_check(): 19 | return JSONResponse(content={"status": "healthy"}, media_type=MediaType.APPLICATION_JSON) -------------------------------------------------------------------------------- /cuda-docker-entrypoint.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | readarray -t gpu_free_mem < <(nvidia-smi --query-gpu=memory.free --format=csv,noheader,nounits) 3 | 4 | if [ "${#gpu_free_mem[@]}" -eq 0 ]; then 5 | echo "Error: Could not retrieve GPU memory information from nvidia-smi." 6 | exit 1 7 | fi 8 | 9 | max_free=-1 10 | max_idx=-1 11 | 12 | for i in "${!gpu_free_mem[@]}"; do 13 | mem="${gpu_free_mem[$i]}" 14 | if [ "$mem" -gt "$max_free" ]; then 15 | max_free="$mem" 16 | max_idx="$i" 17 | fi 18 | done 19 | 20 | export CUDA_VISIBLE_DEVICES="$max_idx" 21 | echo "GPU with index $max_idx has the most available memory (${max_free} MiB)." 22 | echo "Setting CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES" 23 | 24 | exec "$@" -------------------------------------------------------------------------------- /Dockerfile.cpu: -------------------------------------------------------------------------------- 1 | ARG PYTHON_VERSION=3.10 2 | 3 | FROM python:${PYTHON_VERSION}-slim-bookworm AS base 4 | 5 | ENV DEBIAN_FRONTEND=noninteractive 6 | 7 | RUN apt-get update && apt-get install -y --no-install-recommends \ 8 | build-essential \ 9 | curl \ 10 | ffmpeg \ 11 | git \ 12 | && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/* 13 | 14 | WORKDIR /workspace 15 | 16 | FROM base AS python-env 17 | 18 | RUN python3 -m venv /workspace/venv 19 | 20 | ENV PATH="/workspace/venv/bin:$PATH" 21 | 22 | COPY requirements-cpu.txt ./ 23 | 24 | RUN pip install --upgrade pip && \ 25 | pip install --no-cache-dir -r requirements-cpu.txt 26 | 27 | COPY requirements.txt constraints.txt ./ 28 | 29 | RUN pip install --no-cache-dir -c constraints.txt -r requirements.txt 30 | 31 | FROM base AS runtime 32 | 33 | COPY --from=python-env /workspace/venv /workspace/venv 34 | 35 | ENV PATH="/workspace/venv/bin:$PATH" 36 | 37 | WORKDIR /workspace 38 | 39 | COPY src/whisperx_api_server ./whisperx_api_server 40 | 41 | ENV UVICORN_HOST=0.0.0.0 42 | ENV UVICORN_PORT=8000 43 | 44 | CMD ["uvicorn", "--factory", "whisperx_api_server.main:create_app"] -------------------------------------------------------------------------------- /src/whisperx_api_server/logger.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import logging.config 3 | 4 | 5 | def setup_logger(log_level: str) -> None: 6 | assert log_level.upper() in { 7 | "DEBUG", 8 | "INFO", 9 | "WARNING", 10 | "ERROR", 11 | "CRITICAL", 12 | }, log_level 13 | 14 | logging_config = { 15 | "version": 1, # required 16 | "disable_existing_loggers": False, 17 | "formatters": { 18 | "default": { 19 | "()": "uvicorn.logging.DefaultFormatter", 20 | "fmt": "%(asctime)s - %(name)s - %(levelname)s - %(message)s", 21 | "use_colors": True 22 | }, 23 | }, 24 | "handlers": { 25 | "default": { 26 | "formatter": "default", 27 | "class": "logging.StreamHandler", 28 | "stream": "ext://sys.stdout", 29 | }, 30 | }, 31 | "loggers": { 32 | "root": { 33 | "level": log_level.upper(), 34 | "handlers": ["default"], 35 | }, 36 | }, 37 | } 38 | 39 | logging.config.dictConfig(logging_config) -------------------------------------------------------------------------------- /compose.yaml: -------------------------------------------------------------------------------- 1 | services: 2 | whisperx-api-server-cuda: 3 | image: whisperx-api-server-cuda 4 | build: 5 | context: . 6 | dockerfile: Dockerfile.cuda 7 | healthcheck: 8 | test: ["CMD-SHELL", "curl --fail http://localhost:8000/healthcheck || exit 1"] 9 | command: uvicorn --factory whisperx_api_server.main:create_app 10 | ports: 11 | - 8000:8000 12 | volumes: 13 | - hugging_face_cache:/root/.cache/huggingface 14 | - torch_cache:/root/.cache/torch 15 | deploy: 16 | resources: 17 | reservations: 18 | devices: 19 | - driver: nvidia 20 | count: 1 21 | capabilities: [gpu] 22 | whisperx-api-server-cpu: 23 | image: whisperx-api-server-cpu 24 | build: 25 | context: . 26 | dockerfile: Dockerfile.cpu 27 | healthcheck: 28 | test: ["CMD-SHELL", "curl --fail http://localhost:8000/healthcheck || exit 1"] 29 | command: uvicorn --factory whisperx_api_server.main:create_app 30 | ports: 31 | - 8000:8000 32 | volumes: 33 | - hugging_face_cache:/root/.cache/huggingface 34 | - torch_cache:/root/.cache/torch 35 | volumes: 36 | hugging_face_cache: 37 | torch_cache: -------------------------------------------------------------------------------- /.github/workflows/docker-build-and-push-cpu.yml: -------------------------------------------------------------------------------- 1 | name: Deploy CPU image to GHCR 2 | 3 | on: 4 | push: 5 | branches: 6 | - main 7 | - 'v[0-9]+.[0-9]+.[0-9]+.[0-9]+' 8 | 9 | jobs: 10 | push-image: 11 | runs-on: ubuntu-latest 12 | steps: 13 | - name: Checkout code 14 | uses: actions/checkout@v4 15 | - name: Set up Docker Buildx 16 | uses: docker/setup-buildx-action@v3 17 | - name: 'Login to GitHub Container Registry' 18 | uses: docker/login-action@v1 19 | with: 20 | registry: ghcr.io 21 | username: ${{github.actor}} 22 | password: ${{secrets.GITHUB_TOKEN}} 23 | - name: Get package version 24 | id: get_version 25 | run: | 26 | EPOCH_TIME=$(date +%s) 27 | COMMIT_HASH=$(git rev-parse --short HEAD) 28 | BRANCH_NAME=$(git rev-parse --abbrev-ref HEAD) 29 | 30 | if [[ "$BRANCH_NAME" =~ ^v[0-9]+\.[0-9]+\.[0-9]+$ ]]; then 31 | VERSION=$BRANCH_NAME 32 | else 33 | VERSION="${EPOCH_TIME}-${COMMIT_HASH}" 34 | fi 35 | 36 | echo "version=$VERSION" >> $GITHUB_OUTPUT 37 | - name: Build and push CPU image 38 | uses: docker/build-push-action@v6 39 | with: 40 | context: . 41 | file: ./Dockerfile.cpu 42 | cache-from: type=gha 43 | cache-to: type=gha,mode=max 44 | push: true 45 | tags: ghcr.io/nyralei/whisperx-api-server:${{ steps.get_version.outputs.version }}-cpu -------------------------------------------------------------------------------- /.github/workflows/docker-build-and-push-cuda.yml: -------------------------------------------------------------------------------- 1 | name: Deploy CUDA image to GHCR 2 | 3 | on: 4 | push: 5 | branches: 6 | - main 7 | - 'v[0-9]+.[0-9]+.[0-9]+' 8 | 9 | jobs: 10 | push-image: 11 | runs-on: ubuntu-latest 12 | steps: 13 | - name: Checkout code 14 | uses: actions/checkout@v4 15 | - name: Set up Docker Buildx 16 | uses: docker/setup-buildx-action@v3 17 | - name: 'Login to GitHub Container Registry' 18 | uses: docker/login-action@v1 19 | with: 20 | registry: ghcr.io 21 | username: ${{github.actor}} 22 | password: ${{secrets.GITHUB_TOKEN}} 23 | - name: Get package version 24 | id: get_version 25 | run: | 26 | EPOCH_TIME=$(date +%s) 27 | COMMIT_HASH=$(git rev-parse --short HEAD) 28 | BRANCH_NAME=$(git rev-parse --abbrev-ref HEAD) 29 | 30 | if [[ "$BRANCH_NAME" =~ ^v[0-9]+\.[0-9]+\.[0-9]+$ ]]; then 31 | VERSION=$BRANCH_NAME 32 | else 33 | VERSION="${EPOCH_TIME}-${COMMIT_HASH}" 34 | fi 35 | 36 | echo "version=$VERSION" >> $GITHUB_OUTPUT 37 | - name: Build and push CUDA image 38 | uses: docker/build-push-action@v6 39 | with: 40 | context: . 41 | file: ./Dockerfile.cuda 42 | cache-from: type=gha 43 | cache-to: type=gha,mode=max 44 | push: true 45 | tags: ghcr.io/nyralei/whisperx-api-server:${{ steps.get_version.outputs.version }}-cuda -------------------------------------------------------------------------------- /Dockerfile.cuda: -------------------------------------------------------------------------------- 1 | ARG UBUNTU_VERSION=22.04 2 | ARG CUDA_VERSION=12.4.1 3 | ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION} 4 | 5 | FROM ${BASE_CUDA_RUN_CONTAINER} AS base 6 | 7 | ENV DEBIAN_FRONTEND=noninteractive 8 | 9 | RUN apt-get update && apt-get install -y --no-install-recommends \ 10 | build-essential \ 11 | ca-certificates \ 12 | curl \ 13 | ffmpeg \ 14 | git \ 15 | libcudnn8 \ 16 | python3 \ 17 | python3-pip \ 18 | python3-venv \ 19 | && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/* 20 | 21 | WORKDIR /workspace 22 | 23 | FROM base AS python-env 24 | 25 | RUN python3 -m venv /workspace/venv 26 | 27 | ENV PATH="/workspace/venv/bin:$PATH" 28 | 29 | COPY requirements-cuda.txt ./ 30 | 31 | RUN pip install --upgrade pip && \ 32 | pip install --no-cache-dir -r requirements-cuda.txt 33 | 34 | COPY requirements.txt constraints.txt ./ 35 | 36 | RUN pip install --no-cache-dir -c constraints.txt -r requirements.txt 37 | 38 | FROM base AS runtime 39 | 40 | COPY --from=python-env /workspace/venv /workspace/venv 41 | 42 | ENV PATH="/workspace/venv/bin:$PATH" 43 | 44 | WORKDIR /workspace 45 | 46 | COPY src/whisperx_api_server ./whisperx_api_server 47 | 48 | ENV UVICORN_HOST=0.0.0.0 49 | ENV UVICORN_PORT=8000 50 | 51 | COPY ./cuda-docker-entrypoint.sh /workspace/docker-entrypoint.sh 52 | RUN chmod +x /workspace/docker-entrypoint.sh 53 | 54 | ENTRYPOINT [ "/workspace/docker-entrypoint.sh" ] 55 | 56 | CMD ["uvicorn", "--factory", "whisperx_api_server.main:create_app"] 57 | -------------------------------------------------------------------------------- /src/whisperx_api_server/dependencies.py: -------------------------------------------------------------------------------- 1 | from functools import lru_cache 2 | from typing import Annotated 3 | import json 4 | import logging 5 | from fastapi import ( 6 | Depends, 7 | HTTPException, 8 | status 9 | ) 10 | from fastapi.security import HTTPAuthorizationCredentials, HTTPBearer 11 | 12 | from whisperx_api_server.config import Config 13 | 14 | @lru_cache 15 | def get_config() -> Config: 16 | return Config() 17 | 18 | ConfigDependency = Annotated[Config, Depends(get_config)] 19 | 20 | security = HTTPBearer() 21 | 22 | logger = logging.getLogger(__name__) 23 | 24 | async def verify_api_key( 25 | config: ConfigDependency, credentials: Annotated[HTTPAuthorizationCredentials, Depends(security)] 26 | ) -> None: 27 | api_keys = {} 28 | 29 | if config.api_keys_file: 30 | try: 31 | with open(config.api_keys_file, 'r') as f: 32 | api_keys = json.load(f) 33 | except (FileNotFoundError, json.JSONDecodeError) as e: 34 | raise HTTPException( 35 | status_code=status.HTTP_500_INTERNAL_SERVER_ERROR, 36 | detail="API keys file error", 37 | ) from e 38 | 39 | client_name = api_keys.get(credentials.credentials) 40 | 41 | if credentials.credentials != config.api_key and client_name is None: 42 | raise HTTPException(status_code=status.HTTP_403_FORBIDDEN, detail="Invalid API Key") 43 | 44 | if client_name: 45 | logger.info(f"Authorized request from client: '{client_name}'") 46 | else: 47 | logger.info("Authorized request using the default API key") 48 | 49 | ApiKeyDependency = Depends(verify_api_key) -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | ## Overview 2 | 3 | WhisperX API Server is a FastAPI-based server designed to transcribe audio files using the Whisper ASR (Automatic Speech Recognition) model based on WhisperX (https://github.com/m-bain/WhisperX) Python library. The API offers an OpenAI-like interface that allows users to upload audio files and receive transcription results in various formats. It supports customizable options such as different models, languages, temperature settings, and more. 4 | 5 | Features 6 | 1. Audio Transcription: Transcribe audio files using the Whisper ASR model. 7 | 2. Model Caching: Load and cache models for reusability and faster performance. 8 | 3. OpenAI-like API, based on https://platform.openai.com/docs/api-reference/audio/createTranscription and https://platform.openai.com/docs/api-reference/audio/createTranslation 9 | 10 | ## API Endpoints 11 | 12 | ### `POST /v1/audio/transcriptions` 13 | https://platform.openai.com/docs/api-reference/audio/createTranscription 14 | 15 | **Parameters**: 16 | - `file`: The audio file to transcribe. 17 | - `model (str)`: The Whisper model to use. Default is `config.whisper.model`. 18 | - `language (str)`: The language for transcription. Default is `config.default_language`. 19 | - `prompt (str)`: Optional transcription prompt. 20 | - `response_format (str)`: The format of the transcription output. Defaults to `json`. 21 | - `temperature (float)`: Temperature setting for transcription. Default is `0.0`. 22 | - `timestamp_granularities (list)`: Granularity of timestamps, either `segment` or `word`. Default is `["segment"]`. 23 | - `stream (bool)`: Enable streaming mode for real-time transcription. (Doesn't work.) 24 | - `hotwords (str)`: Optional hotwords for transcription. 25 | - `suppress_numerals (bool)`: Option to suppress numerals in the transcription. Default is `True`. 26 | - `highlight_words (bool)`: Highlight words in the transcription output for formats like VTT and SRT. 27 | - `align (bool)`: Option to do transcription timings alignment. Default is `True`. 28 | - `diarize (bool)`: Option to diarize the transcription. Default is `False`. 29 | 30 | **Returns**: Transcription results in the specified format. 31 | 32 | ### `POST /v1/audio/translations` 33 | https://platform.openai.com/docs/api-reference/audio/createTranslation 34 | 35 | **Parameters**: 36 | - `file`: The audio file to translate. 37 | - `model (str)`: The Whisper model to use. Default is `config.whisper.model`. 38 | - `prompt (str)`: Optional translation prompt. 39 | - `response_format (str)`: The format of the translation output. Defaults to `json`. 40 | - `temperature (float)`: Temperature setting for translation. Default is `0.0`. 41 | 42 | **Returns**: Translation results in the specified format. 43 | 44 | ### `GET /healthcheck` 45 | Returns the current health status of the API server. 46 | 47 | ### `GET /models/list` 48 | Lists all loaded models currently available on the server. 49 | 50 | ### `POST /models/unload` 51 | Unloads a specific model from memory cache. 52 | 53 | ### `POST /models/load` 54 | Loads a specified model into memory. 55 | 56 | ### Running the API 57 | 58 | **With Docker**: 59 | 60 | For CPU: 61 | ```bash 62 | docker compose build whisperx-api-server-cpu 63 | 64 | docker compose up whisperx-api-server-cpu 65 | ``` 66 | 67 | For CUDA (GPU): 68 | ```bash 69 | docker compose build whisperx-api-server-cuda 70 | 71 | docker compose up whisperx-api-server-cuda 72 | 73 | ``` 74 | 75 | ## Contributing 76 | 77 | Feel free to submit issues, fork the repository, and send pull requests to contribute to the project. 78 | 79 | ## License 80 | 81 | This project is licensed under the GNU GENERAL PUBLIC LICENSE Version 3. See the `LICENSE` file for details. -------------------------------------------------------------------------------- /src/whisperx_api_server/formatters.py: -------------------------------------------------------------------------------- 1 | from whisperx.utils import WriteSRT, WriteVTT, WriteAudacity 2 | from fastapi.responses import JSONResponse, Response 3 | from whisperx_api_server.config import MediaType 4 | 5 | class ListWriter: 6 | """Helper class to store written lines in memory.""" 7 | def __init__(self): 8 | self.lines = [] 9 | 10 | def write(self, text): 11 | self.lines.append(text) 12 | 13 | def get_output(self): 14 | return ''.join(self.lines) 15 | 16 | def flush(self): 17 | pass 18 | 19 | def update_options(kwargs, defaults): 20 | """ 21 | Helper function to update default options with values from kwargs. 22 | 23 | :param kwargs: Keyword arguments from the function call. 24 | :param defaults: Dictionary of default values. 25 | :return: Updated options dictionary. 26 | """ 27 | options = defaults.copy() 28 | options.update({key: kwargs.get(key, value) for key, value in defaults.items()}) 29 | return options 30 | 31 | def handle_whisperx_format(transcript, writer_class, options): 32 | """ 33 | Helper function to handle "srt", "vtt" and "aud" formats using whisperx writers. 34 | 35 | :param transcript: The transcript dictionary. 36 | :param writer_class: The writer class (WriteSRT, WriteVTT or WriteAudacity). 37 | :param options: Options for the writer. 38 | :return: Formatted output as a string. 39 | """ 40 | writer = writer_class(output_dir=None) 41 | output = ListWriter() 42 | 43 | transcript["segments"]["language"] = transcript["language"] 44 | 45 | writer.write_result(transcript["segments"], output, options) 46 | 47 | return output.get_output() 48 | 49 | def format_transcription(transcript, format, **kwargs) -> Response: 50 | """ 51 | Format a transcript into a given format and return a FastAPI Response object. 52 | 53 | :param transcript: The transcript to format, a dictionary with a "segments" key that contains a list of segments with start and end times and text. 54 | :param format: The format to generate the transcript in. Supported formats are "json", "text", "srt", "vtt" and "aud". 55 | :param kwargs: Additional keyword arguments to pass to the formatter. 56 | :return: A FastAPI Response or JSONResponse object with the formatted transcript and appropriate media type. 57 | """ 58 | # Default options, used for formats imported from whisperx.utils 59 | defaults = { 60 | "max_line_width": 1000, 61 | "max_line_count": None, 62 | "highlight_words": kwargs.get("highlight_words", False), 63 | } 64 | options = update_options(kwargs, defaults) 65 | 66 | if format == "json": 67 | response_data = {"text": transcript.get("text", "")} 68 | return JSONResponse(content=response_data, media_type=MediaType.APPLICATION_JSON) 69 | elif format == "verbose_json": 70 | return JSONResponse(content=transcript, media_type=MediaType.APPLICATION_JSON) 71 | elif format == "vtt_json": 72 | transcript["vtt_text"] = handle_whisperx_format(transcript, WriteVTT, options) 73 | return JSONResponse(content=transcript, media_type=MediaType.APPLICATION_JSON) 74 | elif format == "text": 75 | return Response(content=transcript.get("text", ""), media_type=MediaType.TEXT_PLAIN) 76 | elif format == "srt": 77 | content = handle_whisperx_format(transcript, WriteSRT, options) 78 | return Response(content=content, media_type=MediaType.TEXT_PLAIN) 79 | elif format == "vtt": 80 | content = handle_whisperx_format(transcript, WriteVTT, options) 81 | return Response(content=content, media_type=MediaType.TEXT_VTT) 82 | elif format == "aud": 83 | content = handle_whisperx_format(transcript, WriteAudacity, options) 84 | return Response(content=content, media_type=MediaType.TEXT_PLAIN) 85 | else: 86 | raise ValueError(f"Unsupported format: {format}") 87 | -------------------------------------------------------------------------------- /src/whisperx_api_server/main.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import uuid 3 | import asyncio 4 | from fastapi import ( 5 | FastAPI, 6 | Request 7 | ) 8 | from contextlib import asynccontextmanager 9 | from fastapi.middleware.cors import CORSMiddleware 10 | from starlette.middleware.base import BaseHTTPMiddleware 11 | 12 | from whisperx_api_server.dependencies import ApiKeyDependency, get_config 13 | 14 | from whisperx_api_server.logger import setup_logger 15 | 16 | from whisperx_api_server.models import ( 17 | load_model_instance, 18 | load_align_model_cached, 19 | load_diarize_model_cached, 20 | load_transcribe_pipeline_cached, 21 | ) 22 | 23 | from whisperx_api_server.routers.misc import ( 24 | router as misc_router, 25 | ) 26 | 27 | from whisperx_api_server.routers.models import ( 28 | router as models_router, 29 | ) 30 | 31 | from whisperx_api_server.routers.transcriptions import ( 32 | router as transcribe_router, 33 | ) 34 | 35 | class RequestIDMiddleware(BaseHTTPMiddleware): 36 | async def dispatch(self, request: Request, call_next): 37 | request_id = request.headers.get("X-Request-ID", str(uuid.uuid4())) 38 | request.state.request_id = request_id 39 | response = await call_next(request) 40 | response.headers["X-Request-ID"] = request_id 41 | return response 42 | 43 | @asynccontextmanager 44 | async def lifespan(app: FastAPI): 45 | config = get_config() 46 | logger = logging.getLogger(__name__) 47 | 48 | if config.whisper.preload_model is not None: 49 | logger.info(f"Loading model {config.whisper.preload_model}") 50 | model_instance = await load_model_instance(config.whisper.preload_model) 51 | try: 52 | await load_transcribe_pipeline_cached( 53 | whispermodel=model_instance, 54 | language=getattr(config.default_language, "value", config.default_language), 55 | task="transcribe", 56 | ) 57 | except Exception: 58 | logger.exception("Failed to preload transcribe pipeline; will build on first request") 59 | try: 60 | if config.alignment.preload_model is not None: 61 | logger.info(f"Loading model {config.alignment.preload_model}") 62 | await load_align_model_cached(config.alignment.preload_model) 63 | elif config.alignment.whitelist: 64 | for lang in config.alignment.whitelist: 65 | logger.info(f"Loading model {lang}") 66 | await load_align_model_cached(lang) 67 | except Exception: 68 | logger.exception("Failed to preload alignment model(s); will load on demand") 69 | 70 | try: 71 | if config.diarization.preload_model is not None: 72 | logger.info(f"Loading model {config.diarization.preload_model}") 73 | await load_diarize_model_cached(config.diarization.preload_model) 74 | except Exception: 75 | logger.exception("Failed to preload diarization model; will load on demand") 76 | 77 | yield 78 | 79 | def create_app() -> FastAPI: 80 | config = get_config() 81 | setup_logger(config.log_level) 82 | logger = logging.getLogger(__name__) 83 | 84 | logger.debug(f"Config: {config}") 85 | 86 | dependencies = [] 87 | if config.api_key is not None or config.api_keys_file is not None: 88 | dependencies.append(ApiKeyDependency) 89 | 90 | app = FastAPI(lifespan=lifespan) 91 | 92 | # Misc router is for not protected endpoints like healthcheck 93 | app.include_router(misc_router) 94 | 95 | app.include_router(models_router, dependencies=dependencies) 96 | app.include_router(transcribe_router, dependencies=dependencies) 97 | 98 | if config.allow_origins is not None: 99 | app.add_middleware( 100 | CORSMiddleware, 101 | allow_origins=config.allow_origins, 102 | allow_credentials=True, 103 | allow_methods=["*"], 104 | allow_headers=["*"], 105 | ) 106 | 107 | app.add_middleware(RequestIDMiddleware) 108 | 109 | return app -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # IDEs 7 | .vscode/ 8 | .idea/ 9 | 10 | # Custom 11 | models 12 | compose-dev.yaml 13 | Dockerfile.cuda-dev 14 | scripts 15 | whisperx/ 16 | .scannerwork 17 | Dockerfile.test-dev 18 | 19 | # C extensions 20 | *.so 21 | 22 | # Distribution / packaging 23 | .Python 24 | build/ 25 | develop-eggs/ 26 | dist/ 27 | downloads/ 28 | eggs/ 29 | .eggs/ 30 | lib/ 31 | lib64/ 32 | parts/ 33 | sdist/ 34 | var/ 35 | wheels/ 36 | share/python-wheels/ 37 | *.egg-info/ 38 | .installed.cfg 39 | *.egg 40 | MANIFEST 41 | 42 | # PyInstaller 43 | # Usually these files are written by a python script from a template 44 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 45 | *.manifest 46 | *.spec 47 | 48 | # Installer logs 49 | pip-log.txt 50 | pip-delete-this-directory.txt 51 | 52 | # Unit test / coverage reports 53 | htmlcov/ 54 | .tox/ 55 | .nox/ 56 | .coverage 57 | .coverage.* 58 | .cache 59 | nosetests.xml 60 | coverage.xml 61 | *.cover 62 | *.py,cover 63 | .hypothesis/ 64 | .pytest_cache/ 65 | cover/ 66 | 67 | # Translations 68 | *.mo 69 | *.pot 70 | 71 | # Django stuff: 72 | *.log 73 | local_settings.py 74 | db.sqlite3 75 | db.sqlite3-journal 76 | 77 | # Flask stuff: 78 | instance/ 79 | .webassets-cache 80 | 81 | # Scrapy stuff: 82 | .scrapy 83 | 84 | # Sphinx documentation 85 | docs/_build/ 86 | 87 | # PyBuilder 88 | .pybuilder/ 89 | target/ 90 | 91 | # Jupyter Notebook 92 | .ipynb_checkpoints 93 | 94 | # IPython 95 | profile_default/ 96 | ipython_config.py 97 | 98 | # pyenv 99 | # For a library or package, you might want to ignore these files since the code is 100 | # intended to run in multiple environments; otherwise, check them in: 101 | # .python-version 102 | 103 | # pipenv 104 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 105 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 106 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 107 | # install all needed dependencies. 108 | #Pipfile.lock 109 | 110 | # poetry 111 | # Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control. 112 | # This is especially recommended for binary packages to ensure reproducibility, and is more 113 | # commonly ignored for libraries. 114 | # https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control 115 | #poetry.lock 116 | 117 | # pdm 118 | # Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control. 119 | #pdm.lock 120 | # pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it 121 | # in version control. 122 | # https://pdm.fming.dev/latest/usage/project/#working-with-version-control 123 | .pdm.toml 124 | .pdm-python 125 | .pdm-build/ 126 | 127 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm 128 | __pypackages__/ 129 | 130 | # Celery stuff 131 | celerybeat-schedule 132 | celerybeat.pid 133 | 134 | # SageMath parsed files 135 | *.sage.py 136 | 137 | # Environments 138 | .env 139 | .venv 140 | env/ 141 | venv/ 142 | ENV/ 143 | env.bak/ 144 | venv.bak/ 145 | 146 | # Spyder project settings 147 | .spyderproject 148 | .spyproject 149 | 150 | # Rope project settings 151 | .ropeproject 152 | 153 | # mkdocs documentation 154 | /site 155 | 156 | # mypy 157 | .mypy_cache/ 158 | .dmypy.json 159 | dmypy.json 160 | 161 | # Pyre type checker 162 | .pyre/ 163 | 164 | # pytype static type analyzer 165 | .pytype/ 166 | 167 | # Cython debug symbols 168 | cython_debug/ 169 | 170 | # PyCharm 171 | # JetBrains specific template is maintained in a separate JetBrains.gitignore that can 172 | # be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore 173 | # and can be added to the global gitignore or merged into this file. For a more nuclear 174 | # option (not recommended) you can uncomment the following to ignore the entire idea folder. 175 | #.idea/ 176 | -------------------------------------------------------------------------------- /src/whisperx_api_server/config.py: -------------------------------------------------------------------------------- 1 | from enum import Enum 2 | 3 | from pydantic import BaseModel, Field 4 | from pydantic_settings import BaseSettings, SettingsConfigDict 5 | 6 | 7 | class ResponseFormat(str, Enum): 8 | TEXT = "text" 9 | JSON = "json" 10 | VERBOSE_JSON = "verbose_json" 11 | VTT_JSON = "vtt_json" 12 | SRT = "srt" 13 | VTT = "vtt" 14 | AUD = "aud" # Audacity 15 | 16 | class MediaType(str, Enum): 17 | APPLICATION_JSON = "application/json" 18 | TEXT_PLAIN = "text/plain" 19 | TEXT_VTT = "text/vtt" 20 | 21 | class Language(str, Enum): 22 | AF = "af" 23 | AM = "am" 24 | AR = "ar" 25 | AS = "as" 26 | AZ = "az" 27 | BA = "ba" 28 | BE = "be" 29 | BG = "bg" 30 | BN = "bn" 31 | BO = "bo" 32 | BR = "br" 33 | BS = "bs" 34 | CA = "ca" 35 | CS = "cs" 36 | CY = "cy" 37 | DA = "da" 38 | DE = "de" 39 | EL = "el" 40 | EN = "en" 41 | ES = "es" 42 | ET = "et" 43 | EU = "eu" 44 | FA = "fa" 45 | FI = "fi" 46 | FO = "fo" 47 | FR = "fr" 48 | GL = "gl" 49 | GU = "gu" 50 | HA = "ha" 51 | HAW = "haw" 52 | HE = "he" 53 | HI = "hi" 54 | HR = "hr" 55 | HT = "ht" 56 | HU = "hu" 57 | HY = "hy" 58 | ID = "id" 59 | IS = "is" 60 | IT = "it" 61 | JA = "ja" 62 | JW = "jw" 63 | KA = "ka" 64 | KK = "kk" 65 | KM = "km" 66 | KN = "kn" 67 | KO = "ko" 68 | LA = "la" 69 | LB = "lb" 70 | LN = "ln" 71 | LO = "lo" 72 | LT = "lt" 73 | LV = "lv" 74 | MG = "mg" 75 | MI = "mi" 76 | MK = "mk" 77 | ML = "ml" 78 | MN = "mn" 79 | MR = "mr" 80 | MS = "ms" 81 | MT = "mt" 82 | MY = "my" 83 | NE = "ne" 84 | NL = "nl" 85 | NN = "nn" 86 | NO = "no" 87 | OC = "oc" 88 | PA = "pa" 89 | PL = "pl" 90 | PS = "ps" 91 | PT = "pt" 92 | RO = "ro" 93 | RU = "ru" 94 | SA = "sa" 95 | SD = "sd" 96 | SI = "si" 97 | SK = "sk" 98 | SL = "sl" 99 | SN = "sn" 100 | SO = "so" 101 | SQ = "sq" 102 | SR = "sr" 103 | SU = "su" 104 | SV = "sv" 105 | SW = "sw" 106 | TA = "ta" 107 | TE = "te" 108 | TG = "tg" 109 | TH = "th" 110 | TK = "tk" 111 | TL = "tl" 112 | TR = "tr" 113 | TT = "tt" 114 | UK = "uk" 115 | UR = "ur" 116 | UZ = "uz" 117 | VI = "vi" 118 | YI = "yi" 119 | YO = "yo" 120 | YUE = "yue" 121 | ZH = "zh" 122 | 123 | # https://github.com/OpenNMT/CTranslate2/blob/master/docs/quantization.md 124 | class Quantization(str, Enum): 125 | INT8 = "int8" 126 | INT8_FLOAT16 = "int8_float16" 127 | INT8_BFLOAT16 = "int8_bfloat16" 128 | INT8_FLOAT32 = "int8_float32" 129 | INT16 = "int16" 130 | FLOAT16 = "float16" 131 | BFLOAT16 = "bfloat16" 132 | FLOAT32 = "float32" 133 | DEFAULT = "default" 134 | 135 | 136 | class Device(str, Enum): 137 | CPU = "cpu" 138 | CUDA = "cuda" 139 | AUTO = "auto" 140 | 141 | class VadMethod(str, Enum): 142 | SILERO = "silero" 143 | PYANNOTE = "pyannote" 144 | 145 | class WhisperConfig(BaseModel): 146 | """See https://github.com/SYSTRAN/faster-whisper/blob/master/faster_whisper/transcribe.py#L599.""" 147 | 148 | model: str = Field(default="large-v3") 149 | """ 150 | Default Huggingface model to use for transcription. Note, the model must support being ran using CTranslate2. 151 | This model will be used if no model is specified in the request. 152 | 153 | Models created by authors of `faster-whisper` can be found at https://huggingface.co/Systran 154 | You can find other supported models at https://huggingface.co/models?p=2&sort=trending&search=ctranslate2 and https://huggingface.co/models?sort=trending&search=ct2 155 | """ 156 | inference_device: Device = Field(default=Device.AUTO) 157 | device_index: int | list[int] = Field(default=0) 158 | compute_type: Quantization = Field(default=Quantization.DEFAULT) 159 | cpu_threads: int = Field(default=0) 160 | num_workers: int = Field(default=1) 161 | vad_method: VadMethod = Field(default=VadMethod.PYANNOTE) 162 | vad_model: str = Field(default=None) 163 | vad_options: dict = Field(default=None) 164 | cache: bool = Field(default=True) 165 | preload_model: str = Field(default=None) 166 | local_files_only: bool = Field(default=False) 167 | download_root: str = Field(default=None) 168 | 169 | class AlignConfig(BaseModel): 170 | models: dict = Field(default_factory=dict) 171 | whitelist: list = Field(default_factory=list) 172 | cache: bool = Field(default=True) 173 | preload_model: str = Field(default=None) 174 | 175 | class DiarizeConfig(BaseModel): 176 | cache: bool = Field(default=True) 177 | preload_model: str = Field(default=None) 178 | 179 | class Config(BaseSettings): 180 | """ 181 | Configuration for the application. Values can be set via environment variables. 182 | 183 | Pydantic will automatically handle mapping uppercased environment variables to the corresponding fields. 184 | To populate nested, the environment should be prefixed with the nested field name and an underscore. For example, 185 | the environment variable `LOG_LEVEL` will be mapped to `log_level`, `WHISPER__MODEL`(note the double underscore) to `whisper.model`, to set quantization to int8, use `WHISPER__COMPUTE_TYPE=int8`, etc. 186 | """ 187 | 188 | model_config = SettingsConfigDict(env_nested_delimiter="__") 189 | 190 | api_key: str | None = None 191 | 192 | api_keys_file: str | None = None 193 | 194 | log_level: str = "DEBUG" 195 | 196 | host: str = Field(alias="UVICORN_HOST", default="0.0.0.0") 197 | port: int = Field(alias="UVICORN_PORT", default=8000) 198 | allow_origins: list[str] | None = None 199 | 200 | default_language: Language | None = None 201 | 202 | default_response_format: ResponseFormat = ResponseFormat.JSON 203 | 204 | batch_size: int = 12 205 | 206 | whisper: WhisperConfig = WhisperConfig() 207 | 208 | alignment: AlignConfig = AlignConfig() 209 | 210 | diarization: DiarizeConfig = DiarizeConfig() 211 | 212 | cache_cleanup: bool = True 213 | 214 | audio_cleanup: bool = True -------------------------------------------------------------------------------- /src/whisperx_api_server/routers/models.py: -------------------------------------------------------------------------------- 1 | import logging 2 | from fastapi import APIRouter, Form 3 | from fastapi.responses import JSONResponse 4 | from typing import Annotated 5 | from pydantic import AfterValidator 6 | 7 | import whisperx_api_server.transcriber as transcriber 8 | from whisperx_api_server.dependencies import get_config 9 | from whisperx_api_server.config import ( 10 | Language, 11 | MediaType, 12 | ) 13 | from whisperx_api_server.models import ( 14 | load_model_instance, 15 | load_align_model_cached, 16 | load_diarize_model_cached, 17 | model_instances, 18 | align_model_instances, 19 | diarize_model_instances, 20 | unload_model_object, 21 | ) 22 | 23 | logger = logging.getLogger(__name__) 24 | 25 | router = APIRouter() 26 | 27 | def handle_default_openai_model( 28 | model_name: str 29 | ) -> str: 30 | """Adjust the model name if it defaults to 'whisper-1'.""" 31 | config = get_config() 32 | if model_name == "whisper-1": 33 | logger.info(f"{model_name} is not a valid model name. Using {config.whisper.model} instead.") 34 | return config.whisper.model 35 | return model_name 36 | 37 | ModelName = Annotated[str, AfterValidator(handle_default_openai_model)] 38 | 39 | @router.get( 40 | "/models/list", 41 | description="List loaded models", 42 | tags=["models", "transcribe"], 43 | ) 44 | def list_models(): 45 | global model_instances 46 | return JSONResponse(content={"models": list(model_instances.keys())}, media_type=MediaType.APPLICATION_JSON) 47 | 48 | @router.post( 49 | "/models/unload", 50 | description="Unload a model", 51 | tags=["models", "transcribe"], 52 | ) 53 | def unload_model(model: Annotated[ModelName, Form()]): 54 | try: 55 | if model in model_instances: 56 | model_data = model_instances.pop(model, None) 57 | if model_data is not None: 58 | unload_model_object(model_data) 59 | response_data = {"status": "success"} 60 | else: 61 | response_data = {"status": "error", "message": f"Model {model} not found"} 62 | return JSONResponse(content=response_data, media_type=MediaType.APPLICATION_JSON) 63 | except Exception as e: 64 | return JSONResponse(content={"status": "error", "message": str(e)}, media_type=MediaType.APPLICATION_JSON) 65 | 66 | @router.post( 67 | "/models/load", 68 | description="Load a model", 69 | tags=["models", "transcribe"], 70 | ) 71 | async def load_model(model: Annotated[ModelName, Form()]): 72 | try: 73 | await load_model_instance(model) 74 | return JSONResponse(content={"status": "success", "model": model}, media_type=MediaType.APPLICATION_JSON) 75 | except Exception as e: 76 | return JSONResponse(content={"status": "error", "message": str(e)}, media_type=MediaType.APPLICATION_JSON) 77 | 78 | @router.get( 79 | "/align_models/list", 80 | description="List loaded align models", 81 | tags=["models", "align"], 82 | ) 83 | def list_align_models(): 84 | global align_model_instances 85 | return JSONResponse(content={"models": list(align_model_instances.keys())}, media_type=MediaType.APPLICATION_JSON) 86 | 87 | @router.post( 88 | "/align_models/unload", 89 | description="Unload an align model", 90 | tags=["models", "align"], 91 | ) 92 | def unload_align_model(language: Annotated[Language, Form()]): 93 | try: 94 | if language in align_model_instances: 95 | align_model_data = align_model_instances.pop(language, None) 96 | if align_model_data is not None: 97 | unload_model_object(align_model_data.get("model")) 98 | del align_model_data 99 | response_data = {"status": "success"} 100 | else: 101 | response_data = {"status": "error", "message": f"Model with language {language} not found"} 102 | return JSONResponse(content=response_data, media_type=MediaType.APPLICATION_JSON) 103 | except Exception as e: 104 | return JSONResponse(content={"status": "error", "message": str(e)}, media_type=MediaType.APPLICATION_JSON) 105 | 106 | @router.post( 107 | "/align_models/load", 108 | description="Load an align model", 109 | tags=["models", "align"], 110 | ) 111 | async def load_align_model(language: Annotated[Language, Form()]): 112 | try: 113 | await load_align_model_cached(language) 114 | return JSONResponse(content={"status": "success", "model": language}, media_type=MediaType.APPLICATION_JSON) 115 | except Exception as e: 116 | return JSONResponse(content={"status": "error", "message": str(e)}, media_type=MediaType.APPLICATION_JSON) 117 | 118 | @router.get( 119 | "/diarize_models/list", 120 | description="List loaded diarize models", 121 | tags=["models", "diarize"], 122 | ) 123 | def list_diarize_models(): 124 | global diarize_model_instances 125 | return JSONResponse(content={"models": list(diarize_model_instances.keys())}, media_type=MediaType.APPLICATION_JSON) 126 | 127 | @router.post( 128 | "/diarize_models/unload", 129 | description="Unload a diarize model", 130 | tags=["models", "diarize"], 131 | ) 132 | def unload_diarize_model(model: Annotated[ModelName, Form()]): 133 | try: 134 | if model in diarize_model_instances: 135 | diarize_model_data = diarize_model_instances.pop(model, None) 136 | if diarize_model_data is not None: 137 | unload_model_object(diarize_model_data) 138 | response_data = {"status": "success"} 139 | else: 140 | response_data = {"status": "error", "message": f"Model {model} not found"} 141 | return JSONResponse(content=response_data, media_type=MediaType.APPLICATION_JSON) 142 | except Exception as e: 143 | return JSONResponse(content={"status": "error", "message": str(e)}, media_type=MediaType.APPLICATION_JSON) 144 | 145 | @router.post( 146 | "/diarize_models/load", 147 | description="Load a diarize model", 148 | tags=["models", "diarize"], 149 | ) 150 | async def load_diarize_model(model: Annotated[ModelName, Form()]): 151 | try: 152 | await load_diarize_model_cached(model) 153 | return JSONResponse(content={"status": "success", "model": model}, media_type=MediaType.APPLICATION_JSON) 154 | except Exception as e: 155 | return JSONResponse(content={"status": "error", "message": str(e)}, media_type=MediaType.APPLICATION_JSON) -------------------------------------------------------------------------------- /src/whisperx_api_server/transcriber.py: -------------------------------------------------------------------------------- 1 | import contextlib 2 | import os 3 | from whisperx import transcribe as whisperx_transcribe 4 | from whisperx import audio as whisperx_audio 5 | from whisperx import alignment as whisperx_alignment 6 | from whisperx import diarize as whisperx_diarize 7 | from whisperx import types as whisperx_types 8 | from fastapi import UploadFile 9 | import logging 10 | import time 11 | import tempfile 12 | import asyncio 13 | import torch 14 | import gc 15 | 16 | from whisperx_api_server.config import ( 17 | Language, 18 | ) 19 | from whisperx_api_server.dependencies import get_config 20 | from whisperx_api_server.models import ( 21 | CustomWhisperModel, 22 | load_align_model_cached, 23 | load_diarize_model_cached, 24 | load_transcribe_pipeline_cached, 25 | ) 26 | 27 | logger = logging.getLogger(__name__) 28 | 29 | config = get_config() 30 | 31 | _concurrency_semaphore = None 32 | 33 | def _get_concurrency_semaphore() -> asyncio.Semaphore | None: 34 | """Return a semaphore only if running on GPU.""" 35 | global _concurrency_semaphore 36 | if not torch.cuda.is_available(): 37 | return None 38 | if _concurrency_semaphore is None: 39 | max_concurrent = int(os.getenv("MAX_CONCURRENT_TRANSCRIPTIONS", "1")) 40 | _concurrency_semaphore = asyncio.Semaphore(max_concurrent) 41 | return _concurrency_semaphore 42 | 43 | def _cleanup_cache_only(): 44 | gc.collect() 45 | 46 | if torch.cuda.is_available(): 47 | torch.cuda.empty_cache() 48 | 49 | async def _save_upload_to_temp(audio_file: UploadFile, request_id: str) -> str: 50 | loop = asyncio.get_running_loop() 51 | try: 52 | file_bytes = await audio_file.read() 53 | except Exception as e: 54 | logger.error(f"Request ID: {request_id} - Failed to read uploaded file: {e}") 55 | raise 56 | 57 | def _write_temp_file(data: bytes) -> str: 58 | with tempfile.NamedTemporaryFile(delete=False, suffix=f"_{audio_file.filename}") as temp_file: 59 | temp_file.write(data) 60 | return temp_file.name 61 | 62 | try: 63 | file_path = await loop.run_in_executor(None, _write_temp_file, file_bytes) 64 | except Exception as e: 65 | logger.error(f"Request ID: {request_id} - Failed to write temp file: {e}") 66 | raise 67 | 68 | return file_path 69 | 70 | 71 | async def _load_audio(file_path: str, request_id: str): 72 | loop = asyncio.get_running_loop() 73 | try: 74 | audio = await loop.run_in_executor(None, whisperx_audio.load_audio, file_path) 75 | logger.info(f"Request ID: {request_id} - Audio loaded from {file_path}") 76 | return audio 77 | except Exception as e: 78 | logger.error(f"Request ID: {request_id} - Failed to load audio: {e}") 79 | raise 80 | 81 | async def _transcribe_audio(model, audio, batch_size, chunk_size, language, task, request_id): 82 | loop = asyncio.get_running_loop() 83 | 84 | def _run_transcription(): 85 | with torch.inference_mode(): 86 | return model.transcribe( 87 | audio=audio, 88 | batch_size=batch_size, 89 | chunk_size=chunk_size, 90 | num_workers=config.whisper.num_workers, 91 | language=language, 92 | task=task, 93 | ) 94 | 95 | result = await loop.run_in_executor(None, _run_transcription) 96 | 97 | logger.info(f"Request ID: {request_id} - Transcription completed") 98 | return result 99 | 100 | 101 | async def _align_audio(result, audio, whispermodel, request_id): 102 | loop = asyncio.get_running_loop() 103 | try: 104 | alignment_model_start = time.time() 105 | logger.info(f"Request ID: {request_id} - Loading alignment model") 106 | model_a, metadata = await load_align_model_cached(language_code=result["language"]) 107 | logger.info(f"Request ID: {request_id} - Alignment model loaded") 108 | logger.info(f"Request ID: {request_id} - Loading alignment model took {time.time() - alignment_model_start:.2f} seconds") 109 | 110 | def _run_alignment(): 111 | with torch.inference_mode(): 112 | return whisperx_alignment.align( 113 | transcript=result["segments"], 114 | model=model_a, 115 | align_model_metadata=metadata, 116 | audio=audio, 117 | device=whispermodel.device, 118 | return_char_alignments=False 119 | ) 120 | alignment_start = time.time() 121 | result["segments"] = await loop.run_in_executor(None, _run_alignment) 122 | logger.info(f"Request ID: {request_id} - Alignment took {time.time() - alignment_start:.2f} seconds") 123 | return result 124 | except Exception as e: 125 | logger.error(f"Request ID: {request_id} - Alignment failed: {e}") 126 | raise 127 | 128 | 129 | async def _diarize_audio(result, audio, request_id): 130 | loop = asyncio.get_running_loop() 131 | try: 132 | diarization_model_start = time.time() 133 | logger.info(f"Request ID: {request_id} - Loading diarization model") 134 | diarize_model = await load_diarize_model_cached(model_name="tensorlake/speaker-diarization-3.1") 135 | logger.info(f"Request ID: {request_id} - Diarization model loaded. Loading took {time.time() - diarization_model_start:.2f} seconds. Starting diarization") 136 | 137 | def _run_diarization(): 138 | with torch.inference_mode(): 139 | return diarize_model(audio) 140 | diarize_start = time.time() 141 | diarize_segments = await loop.run_in_executor(None, _run_diarization) 142 | result["segments"] = whisperx_diarize.assign_word_speakers(diarize_segments, result["segments"]) 143 | logger.info(f"Request ID: {request_id} - Diarization took {time.time() - diarize_start:.2f} seconds") 144 | return result 145 | except Exception as e: 146 | logger.error(f"Request ID: {request_id} - Diarization failed: {e}") 147 | raise 148 | 149 | def _finalize_text(result, align_or_diarize: bool): 150 | segments = result.get("segments", []) 151 | if align_or_diarize and isinstance(segments, dict): 152 | segments = segments.get("segments", []) 153 | 154 | result["text"] = '\n'.join([s.get("text", "").strip() for s in segments if s.get("text")]) 155 | return result 156 | 157 | async def transcribe( 158 | audio_file: UploadFile, 159 | batch_size: int = config.batch_size, 160 | chunk_size: int = 30, 161 | asr_options: dict = {}, 162 | language: Language = config.default_language, 163 | whispermodel: CustomWhisperModel = config.whisper.model, 164 | align: bool = False, 165 | diarize: bool = False, 166 | request_id: str = "", 167 | task: str = "transcribe", 168 | ) -> whisperx_types.TranscriptionResult: 169 | start_time = time.time() 170 | file_path = None 171 | audio = None 172 | concurrency_sem = _get_concurrency_semaphore() 173 | 174 | try: 175 | file_path = await _save_upload_to_temp(audio_file, request_id) 176 | logger.info(f"Request ID: {request_id} - Saving uploaded file took {time.time() - start_time:.2f} seconds") 177 | 178 | if concurrency_sem: 179 | await concurrency_sem.acquire() 180 | logger.debug(f"Request ID: {request_id} - Acquired GPU concurrency semaphore") 181 | 182 | logger.info(f"Request ID: {request_id} - Transcribing {audio_file.filename} with model: {whispermodel.model_size_or_path} and options: {asr_options}, language: {language}, task: {task}") 183 | 184 | model_loading_start = time.time() 185 | 186 | model = await load_transcribe_pipeline_cached( 187 | whispermodel=whispermodel, 188 | language=language, 189 | task=task, 190 | ) 191 | 192 | logger.info(f"Request ID: {request_id} - Loading model took {time.time() - model_loading_start:.2f} seconds (cached)") 193 | 194 | audio_loading_start = time.time() 195 | 196 | audio = await _load_audio(file_path, request_id) 197 | 198 | logger.info(f"Request ID: {request_id} - Loading audio took {time.time() - audio_loading_start:.2f} seconds") 199 | 200 | transcription_start = time.time() 201 | 202 | result = await _transcribe_audio(model, audio, batch_size, chunk_size, language, task, request_id) 203 | 204 | logger.info(f"Request ID: {request_id} - Transcription took {time.time() - transcription_start:.2f} seconds") 205 | 206 | if align or diarize: 207 | result = await _align_audio(result, audio, whispermodel, request_id) 208 | 209 | if diarize: 210 | result = await _diarize_audio(result, audio, request_id) 211 | 212 | result = _finalize_text(result, align or diarize) 213 | 214 | logger.info(f"Request ID: {request_id} - Transcription completed for {audio_file.filename}") 215 | 216 | return result 217 | except Exception as e: 218 | logger.error(f"Request ID: {request_id} - Transcription failed for {audio_file.filename} with error: {e}") 219 | raise 220 | finally: 221 | with contextlib.suppress(Exception): 222 | if concurrency_sem: 223 | concurrency_sem.release() 224 | with contextlib.suppress(Exception): 225 | if file_path and os.path.exists(file_path): 226 | os.remove(file_path) 227 | if config.audio_cleanup and audio is not None: 228 | del audio 229 | logger.info(f"Request ID: {request_id} - Audio data cleaned up") 230 | if config.cache_cleanup: 231 | _cleanup_cache_only() 232 | logger.info(f"Request ID: {request_id} - Cache cleanup completed") -------------------------------------------------------------------------------- /src/whisperx_api_server/routers/transcriptions.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import uuid 3 | from .models import handle_default_openai_model 4 | from fastapi import ( 5 | APIRouter, 6 | UploadFile, 7 | Form, 8 | HTTPException, 9 | Request, 10 | status 11 | ) 12 | from fastapi.responses import Response 13 | from starlette.middleware.base import BaseHTTPMiddleware 14 | from typing import Literal, Annotated 15 | from pydantic import AfterValidator 16 | import time 17 | 18 | 19 | import whisperx_api_server.transcriber as transcriber 20 | from whisperx_api_server.dependencies import ConfigDependency 21 | from whisperx_api_server.formatters import format_transcription 22 | from whisperx_api_server.config import ( 23 | Language, 24 | ResponseFormat, 25 | ) 26 | from whisperx_api_server.models import ( 27 | load_model_instance, 28 | ) 29 | 30 | logger = logging.getLogger(__name__) 31 | 32 | router = APIRouter() 33 | 34 | # Annotated ModelName for validation and defaults 35 | ModelName = Annotated[str, AfterValidator(handle_default_openai_model)] 36 | 37 | class RequestIDMiddleware(BaseHTTPMiddleware): 38 | async def dispatch(self, request: Request, call_next): 39 | request_id = request.headers.get("X-Request-ID", str(uuid.uuid4())) 40 | request.state.request_id = request_id 41 | response = await call_next(request) 42 | response.headers["X-Request-ID"] = request_id 43 | return response 44 | 45 | async def get_timestamp_granularities(request: Request) -> list[Literal["segment", "word"]]: 46 | TIMESTAMP_GRANULARITIES_COMBINATIONS = [ 47 | [], 48 | ["segment"], 49 | ["word"], 50 | ["word", "segment"], 51 | ["segment", "word"], 52 | ] 53 | form = await request.form() 54 | if form.get("timestamp_granularities[]") is None: 55 | return ["segment"] 56 | timestamp_granularities = form.getlist("timestamp_granularities[]") 57 | assert timestamp_granularities in TIMESTAMP_GRANULARITIES_COMBINATIONS, ( 58 | f"{timestamp_granularities} is not a valid value for `timestamp_granularities[]`." 59 | ) 60 | return timestamp_granularities 61 | 62 | def apply_defaults(config, model, language=None, response_format=None): 63 | if model is None: 64 | model = config.whisper.model 65 | if language is None: 66 | language = config.default_language 67 | if response_format is None: 68 | response_format = config.default_response_format 69 | return model, language, response_format 70 | 71 | """ 72 | OpenAI-like endpoint to transcribe audio files using the Whisper ASR model. 73 | 74 | Args: 75 | request (Request): The HTTP request object. 76 | file (UploadFile): The audio file to transcribe. 77 | model (ModelName): The model to use for the transcription. 78 | language (Language): The language to use for the transcription. Defaults to "en". 79 | prompt (str): The prompt to use for the transcription. 80 | response_format (ResponseFormat): The response format to use for the transcription. Defaults to "json". 81 | temperature (float): The temperature to use for the transcription. Defaults to 0.0. 82 | timestamp_granularities (list[Literal["segment", "word"]]): The timestamp granularities to use for the transcription. Defaults to ["segment"]. 83 | stream (bool): Whether to enable streaming mode. Defaults to False. 84 | hotwords (str): The hotwords to use for the transcription. 85 | suppress_numerals (bool): Whether to suppress numerals in the transcription. Defaults to True. 86 | highlight_words (bool): Whether to highlight words in the transcription (Applies only to VTT and SRT). Defaults to False. 87 | align (bool): Whether to do transcription timings alignment. Defaults to True. 88 | diarize (bool): Whether to diarize the transcription. Defaults to False. 89 | chunk_size (int): Chunk size in seconds for merging VAD segments. Defaults to 30. 90 | 91 | Returns: 92 | Transcription: The transcription of the audio file. 93 | """ 94 | @router.post( 95 | "/v1/audio/transcriptions", 96 | description="Transcribe audio files using the Whisper ASR model.", 97 | tags=["Transcription"], 98 | ) 99 | async def transcribe_audio( 100 | config: ConfigDependency, 101 | request: Request, 102 | file: UploadFile, 103 | model: Annotated[ModelName, Form()] = None, 104 | language: Annotated[Language, Form()] = None, 105 | prompt: Annotated[str, Form()] = None, 106 | response_format: Annotated[ResponseFormat, Form()] = None, 107 | temperature: Annotated[float, Form()] = 0.0, 108 | timestamp_granularities: Annotated[ 109 | list[Literal["segment", "word"]], 110 | Form(alias="timestamp_granularities[]"), 111 | ] = ["segment"], 112 | stream: Annotated[bool, Form()] = False, 113 | hotwords: Annotated[str, Form()] = None, 114 | suppress_numerals: Annotated[bool, Form()] = True, 115 | highlight_words: Annotated[bool, Form()] = False, 116 | align: Annotated[bool, Form()] = True, 117 | diarize: Annotated[bool, Form()] = False, 118 | chunk_size: Annotated[int, Form()] = 30, 119 | ) -> Response: 120 | model, language, response_format = apply_defaults(config, model, language, response_format) 121 | timestamp_granularities = await get_timestamp_granularities(request) 122 | request_id = request.state.request_id 123 | logger.info(f"Request ID: {request_id} - Received transcription request") 124 | start_time = time.time() # Start the timer 125 | logger.info(f"Request ID: {request_id} - Received request to transcribe {file.filename} with parameters: \ 126 | model: {model}, \ 127 | language: {language}, \ 128 | prompt: {prompt}, \ 129 | response_format: {response_format}, \ 130 | temperature: {temperature}, \ 131 | timestamp_granularities: {timestamp_granularities}, \ 132 | stream: {stream}, \ 133 | hotwords: {hotwords}, \ 134 | suppress_numerals: {suppress_numerals}, \ 135 | highlight_words: {highlight_words}, \ 136 | align: {align}, \ 137 | diarize: {diarize}, \ 138 | chunk_size: {chunk_size}") 139 | 140 | if not align: 141 | if response_format in ('vtt', 'srt', 'aud', 'vtt_json'): 142 | raise HTTPException( 143 | status_code=status.HTTP_500_INTERNAL_SERVER_ERROR, 144 | detail="Subtitles format ('vtt', 'srt', 'aud', 'vtt_json') requires alignment to be enabled." 145 | ) 146 | 147 | if diarize: 148 | raise HTTPException( 149 | status_code=status.HTTP_500_INTERNAL_SERVER_ERROR, 150 | detail="Diarization requires alignment to be enabled." 151 | ) 152 | 153 | # Determine if word timestamps are required 154 | word_timestamps = "word" in timestamp_granularities 155 | 156 | # Build ASR options 157 | asr_options = { 158 | "suppress_numerals": suppress_numerals, 159 | "temperatures": temperature, 160 | "word_timestamps": word_timestamps, 161 | "initial_prompt": prompt, 162 | "hotwords": hotwords, 163 | } 164 | 165 | model_load_time = time.time() 166 | # Get model instance (reuse if cached) 167 | model_instance = await load_model_instance(model) 168 | 169 | logger.info(f"Loaded model {model} in {time.time() - model_load_time:.2f} seconds") 170 | 171 | try: 172 | transcription = await transcriber.transcribe( 173 | audio_file=file, 174 | asr_options=asr_options, 175 | language=language, 176 | whispermodel=model_instance, 177 | align=align, 178 | diarize=diarize, 179 | chunk_size=chunk_size, 180 | request_id=request_id 181 | ) 182 | except Exception as e: 183 | logger.exception(f"Request ID: {request_id} - Transcription failed: {e}") 184 | raise HTTPException( 185 | status_code=status.HTTP_500_INTERNAL_SERVER_ERROR, 186 | detail="An unexpected error occurred while processing the transcription request." 187 | ) from e 188 | 189 | total_time = time.time() - start_time 190 | logger.info(f"Request ID: {request_id} - Transcription process took {total_time:.2f} seconds") 191 | 192 | return format_transcription(transcription, response_format, highlight_words=highlight_words) 193 | 194 | """ 195 | OpenAI-like endpoint to translate audio files using the Whisper ASR model. 196 | 197 | Args: 198 | request (Request): The HTTP request object. 199 | file (UploadFile): The audio file to translate. 200 | model (ModelName): The model to use for the translation. 201 | prompt (str): The prompt to use for the translation. 202 | response_format (ResponseFormat): The response format to use for the translation. Defaults to "json". 203 | temperature (float): The temperature to use for the translation. Defaults to 0.0. 204 | chunk_size (int): Chunk size in seconds for merging VAD segments. Defaults to 30. 205 | 206 | Returns: 207 | Translation: The translation of the audio file. 208 | """ 209 | @router.post( 210 | "/v1/audio/translations", 211 | description="Translate audio files using the Whisper ASR model", 212 | tags=["Translation"], 213 | ) 214 | async def translate_audio( 215 | config: ConfigDependency, 216 | request: Request, 217 | file: UploadFile, 218 | model: Annotated[ModelName, Form()] = None, 219 | prompt: Annotated[str, Form()] = "", 220 | response_format: Annotated[ResponseFormat, Form()] = None, 221 | temperature: Annotated[float, Form()] = 0.0, 222 | chunk_size: Annotated[int, Form()] = 30, 223 | ) -> Response: 224 | model, _, response_format = apply_defaults(config, model, language=None, response_format=response_format) 225 | request_id = request.state.request_id 226 | logger.info(f"Request ID: {request_id} - Received translation request") 227 | start_time = time.time() # Start the timer 228 | logger.info(f"Request ID: {request_id} - Received request to translate {file.filename} with parameters: \ 229 | model: {model}, \ 230 | prompt: {prompt}, \ 231 | response_format: {response_format}, \ 232 | temperature: {temperature}, \ 233 | chunk_size: {chunk_size}") 234 | 235 | # Build ASR options 236 | asr_options = { 237 | "initial_prompt": prompt, 238 | "temperatures": temperature, 239 | } 240 | 241 | model_load_time = time.time() 242 | # Get model instance (reuse if cached) 243 | model_instance = await load_model_instance(model) 244 | 245 | logger.info(f"Loaded model {model} in {time.time() - model_load_time:.2f} seconds") 246 | 247 | try: 248 | translation = await transcriber.transcribe( 249 | audio_file=file, 250 | asr_options=asr_options, 251 | whispermodel=model_instance, 252 | chunk_size=chunk_size, 253 | request_id=request_id, 254 | task="translate" 255 | ) 256 | except Exception as e: 257 | logger.exception(f"Request ID: {request_id} - Translation failed: {e}") 258 | raise HTTPException( 259 | status_code=status.HTTP_500_INTERNAL_SERVER_ERROR, 260 | detail="An unexpected error occurred while processing the translation request." 261 | ) from e 262 | 263 | total_time = time.time() - start_time 264 | logger.info(f"Request ID: {request_id} - Translation process took {total_time:.2f} seconds") 265 | 266 | return format_transcription(translation, response_format) -------------------------------------------------------------------------------- /src/whisperx_api_server/models.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import asyncio 3 | import contextlib 4 | import torch 5 | import gc 6 | from collections import defaultdict 7 | from asyncio import Lock 8 | from typing import Union, List, Optional, Tuple, Any 9 | 10 | from whisperx import asr as whisperx_asr 11 | from whisperx import transcribe as whisperx_transcribe 12 | from whisperx import alignment as whisperx_alignment 13 | from whisperx import diarize as whisperx_diarize 14 | 15 | from whisperx_api_server.dependencies import get_config 16 | 17 | logger = logging.getLogger(__name__) 18 | 19 | # Global caches 20 | model_instances = {} 21 | model_locks = defaultdict(Lock) 22 | 23 | align_model_instances = {} 24 | alignment_locks = defaultdict(Lock) 25 | 26 | diarize_model_instances = {} 27 | diarization_locks = defaultdict(Lock) 28 | 29 | alignment_cache_mod_lock = Lock() 30 | 31 | # Transcribe pipeline cache 32 | transcribe_pipeline_instances = {} 33 | transcribe_locks = defaultdict(Lock) 34 | 35 | def unload_model_object(model_obj: Any): 36 | if model_obj is None: 37 | return 38 | # 1) Move to CPU 39 | with contextlib.suppress(Exception): 40 | model_obj.to("cpu") 41 | # 2) Delete reference 42 | del model_obj 43 | 44 | # 3) Force GC and empty cache 45 | gc.collect() 46 | torch.cuda.empty_cache() 47 | 48 | class CustomWhisperModel(whisperx_asr.WhisperModel): 49 | def __init__( 50 | self, 51 | model_size_or_path: str, 52 | device: str = "auto", 53 | device_index: Union[int, List[int]] = 0, 54 | compute_type: str = "default", 55 | cpu_threads: int = 0, 56 | num_workers: int = 1, 57 | download_root: Optional[str] = None, 58 | local_files_only: bool = False, 59 | files: dict = None, 60 | **model_kwargs, 61 | ): 62 | super().__init__( 63 | model_size_or_path=model_size_or_path, 64 | device=device, 65 | device_index=device_index, 66 | compute_type=compute_type, 67 | cpu_threads=cpu_threads, 68 | num_workers=num_workers, 69 | download_root=download_root, 70 | local_files_only=local_files_only, 71 | files=files, 72 | **model_kwargs 73 | ) 74 | self.model_size_or_path = model_size_or_path 75 | self.device = device 76 | self.device_index = device_index 77 | self.compute_type = compute_type 78 | self.cpu_threads = cpu_threads 79 | self.num_workers = num_workers 80 | self.download_root = download_root 81 | self.local_files_only = local_files_only 82 | self.files = files 83 | self.model_kwargs = model_kwargs 84 | 85 | 86 | def check_device(): 87 | """ 88 | Original device-checking function used by whisper. 89 | """ 90 | try: 91 | return "cuda" if torch.cuda.is_available() else "cpu" 92 | except Exception: 93 | logger.error("Could not determine device. Using 'cpu' instead.") 94 | return "cpu" 95 | 96 | 97 | def _determine_inference_device(): 98 | """ 99 | Helper to fetch `inference_device` from config or default to CPU/GPU auto-check. 100 | """ 101 | config = get_config() 102 | inference_device = config.whisper.inference_device.value 103 | if inference_device == "auto": 104 | inference_device = check_device() 105 | return inference_device 106 | 107 | 108 | def initialize_model(model_name: str) -> CustomWhisperModel: 109 | """ 110 | Initializes a CustomWhisperModel with the config settings. 111 | """ 112 | config = get_config() 113 | inference_device = _determine_inference_device() 114 | return CustomWhisperModel( 115 | model_size_or_path=model_name, 116 | device=inference_device, 117 | device_index=config.whisper.device_index, 118 | compute_type=config.whisper.compute_type.value, 119 | cpu_threads=config.whisper.cpu_threads, 120 | num_workers=config.whisper.num_workers, 121 | local_files_only=config.whisper.local_files_only, 122 | download_root=config.whisper.download_root 123 | ) 124 | 125 | 126 | async def _get_or_init_model( 127 | key: str, 128 | cache_dict: dict, 129 | lock_dict: dict, 130 | init_func, 131 | log_reuse: str = "Reusing cached model instance for {key}", 132 | log_init: str = "Initializing model: {key}", 133 | ) -> Any: 134 | """ 135 | Generic helper: 136 | 1) Check if `key` is in `cache_dict`. 137 | 2) If not, acquire lock_dict[key] and load the model. 138 | 3) Return the cached/loaded model. 139 | """ 140 | if key in cache_dict: 141 | logger.info(log_reuse.format(key=key)) 142 | return cache_dict[key] 143 | 144 | async with lock_dict[key]: 145 | # Double-check after acquiring the lock 146 | if key not in cache_dict: 147 | logger.info(log_init.format(key=key)) 148 | cache_dict[key] = await init_func() 149 | return cache_dict[key] 150 | 151 | 152 | # ------------------------------------------------------------------------- 153 | # Main Whisper model loading 154 | # ------------------------------------------------------------------------- 155 | async def load_model_instance(model_name: str): 156 | """ 157 | Async function to get the main Whisper model instance from cache, or initialize if needed. 158 | """ 159 | return await _get_or_init_model( 160 | key=model_name, 161 | cache_dict=model_instances, 162 | lock_dict=model_locks, 163 | init_func=lambda: asyncio.to_thread(initialize_model, model_name), 164 | ) 165 | 166 | # ------------------------------------------------------------------------- 167 | # Transcribe pipeline loading 168 | # ------------------------------------------------------------------------- 169 | def _hashable_vad_options(vad_options: Any) -> Any: 170 | if vad_options is None: 171 | return None 172 | if isinstance(vad_options, dict): 173 | return tuple(sorted((k, _hashable_vad_options(v)) for k, v in vad_options.items())) 174 | if isinstance(vad_options, (list, tuple)): 175 | return tuple(_hashable_vad_options(v) for v in vad_options) 176 | return vad_options 177 | 178 | async def load_transcribe_pipeline_cached( 179 | whispermodel: CustomWhisperModel, 180 | language: Optional[str] = None, 181 | task: str = "transcribe", 182 | ): 183 | config = get_config() 184 | key = ( 185 | whispermodel.model_size_or_path, 186 | whispermodel.device, 187 | whispermodel.compute_type, 188 | config.whisper.vad_method.value if hasattr(config.whisper.vad_method, "value") else config.whisper.vad_method, 189 | config.whisper.vad_model, 190 | _hashable_vad_options(config.whisper.vad_options), 191 | ) 192 | 193 | def _init_pipeline(): 194 | return whisperx_transcribe.load_model( 195 | whisper_arch=whispermodel.model_size_or_path, 196 | device=whispermodel.device, 197 | compute_type=whispermodel.compute_type, 198 | language=language, 199 | vad_model=config.whisper.vad_model, 200 | vad_method=config.whisper.vad_method, 201 | vad_options=config.whisper.vad_options, 202 | task=task, 203 | ) 204 | 205 | pipeline = await _get_or_init_model( 206 | key=str(key), 207 | cache_dict=transcribe_pipeline_instances, 208 | lock_dict=transcribe_locks, 209 | init_func=lambda: asyncio.to_thread(_init_pipeline), 210 | log_reuse="Reusing cached transcribe pipeline: {key}", 211 | log_init="Initializing transcribe pipeline: {key}", 212 | ) 213 | 214 | if not config.whisper.cache: 215 | removed = transcribe_pipeline_instances.pop(str(key), None) 216 | if removed is not None: 217 | logger.info(f"Unloading transcribe pipeline from cache (disabled): {key}") 218 | unload_model_object(removed) 219 | 220 | return pipeline 221 | 222 | # ------------------------------------------------------------------------- 223 | # Alignment model loading 224 | # ------------------------------------------------------------------------- 225 | async def _cleanup_alignment_cache_whitelist(): 226 | """ 227 | If config.alignment.whitelist is set, remove any alignment models from 228 | `align_model_instances` that are not in the whitelist. 229 | This happens under a dedicated lock to avoid race conditions. 230 | """ 231 | config = get_config() 232 | whitelist = config.alignment.whitelist 233 | if not whitelist: 234 | return 235 | 236 | async with alignment_cache_mod_lock: 237 | for key in list(align_model_instances.keys()): # noqa: S7504 238 | if key not in whitelist: 239 | logger.info(f"Unloading alignment model for {key} (not in whitelist).") 240 | align_model_data = align_model_instances.pop(key, None) 241 | if align_model_data is not None: 242 | unload_model_object(align_model_data.get("model")) 243 | del align_model_data 244 | 245 | async def load_align_model_cached( 246 | language_code: str, 247 | model_name: Optional[str] = None, 248 | model_dir: Optional[str] = None 249 | ) -> Tuple[Any, Any]: 250 | """ 251 | Loads and caches alignment models based on language codes (or "multilingual") 252 | while respecting the config whitelisting and caching settings. 253 | """ 254 | config = get_config() 255 | 256 | # Clean up out-of-whitelist models 257 | await _cleanup_alignment_cache_whitelist() 258 | 259 | inference_device = _determine_inference_device() 260 | 261 | selected_model_name = model_name 262 | if "multilingual" in config.alignment.models: 263 | selected_model_name = config.alignment.models["multilingual"] 264 | logger.info(f"Overriding with 'multilingual' alignment model: {selected_model_name}") 265 | elif language_code in config.alignment.models: 266 | selected_model_name = config.alignment.models[language_code] 267 | logger.info(f"Using configured alignment model for '{language_code}': {selected_model_name}") 268 | 269 | # Decide how to key the cache 270 | if (selected_model_name is not None 271 | and selected_model_name == config.alignment.models.get("multilingual")): 272 | cache_key = "multilingual" 273 | else: 274 | cache_key = language_code 275 | 276 | logger.debug(f"config.alignment.models = {config.alignment.models}") 277 | logger.debug(f"Incoming language_code = {language_code}, model_name param = {model_name}") 278 | 279 | async def _init_alignment(): 280 | try: 281 | loop = asyncio.get_running_loop() 282 | align_model, align_metadata = await loop.run_in_executor( 283 | None, # or a custom ThreadPoolExecutor 284 | lambda: whisperx_alignment.load_align_model( 285 | language_code=language_code, 286 | device=inference_device, 287 | model_name=selected_model_name, 288 | model_dir=model_dir 289 | ) 290 | ) 291 | except Exception as e: 292 | logger.error(f"Failed to load alignment model for language '{language_code}': {e}") 293 | raise 294 | 295 | return {"model": align_model, "metadata": align_metadata} 296 | 297 | # Fetch or initialize the alignment model under a lock: 298 | model_data = await _get_or_init_model( 299 | key=cache_key, 300 | cache_dict=align_model_instances, 301 | lock_dict=alignment_locks, 302 | init_func=_init_alignment, 303 | log_reuse="Reusing cached alignment model for key: {key}", 304 | log_init="Initializing alignment model for key: {key}", 305 | ) 306 | 307 | # If caching is disabled, remove it immediately and free GPU memory 308 | if not config.alignment.cache: 309 | async with alignment_cache_mod_lock: 310 | removed_data = align_model_instances.pop(cache_key, None) 311 | if removed_data is not None: 312 | logger.info(f"Unloading alignment model from cache (disabled): {cache_key}") 313 | model_obj = removed_data.get("model") 314 | if model_obj is not None: 315 | unload_model_object(model_obj) 316 | del removed_data 317 | 318 | return model_data["model"], model_data["metadata"] 319 | 320 | 321 | # ------------------------------------------------------------------------- 322 | # Diarization model loading 323 | # ------------------------------------------------------------------------- 324 | async def load_diarize_model_cached(model_name: str): 325 | """ 326 | Loads and caches a diarization pipeline model, if not already present. 327 | Clears from cache after use if `config.diarization.cache` is False. 328 | """ 329 | config = get_config() 330 | inference_device = _determine_inference_device() 331 | 332 | def _init_diarization(): 333 | logger.info(f"Loading diarization pipeline for model: {model_name} with device: {inference_device}") 334 | return whisperx_diarize.DiarizationPipeline(model_name=model_name, device=inference_device) 335 | 336 | diarize_model = await _get_or_init_model( 337 | key=model_name, 338 | cache_dict=diarize_model_instances, 339 | lock_dict=diarization_locks, 340 | init_func=lambda: asyncio.to_thread(_init_diarization), 341 | log_reuse="Reusing cached diarization model for: {key}", 342 | log_init="Initializing diarization model: {key}", 343 | ) 344 | 345 | if not config.diarization.cache: 346 | # Immediately remove from cache, unload from GPU memory 347 | removed_model = diarize_model_instances.pop(model_name, None) 348 | if removed_model is not None: 349 | logger.info(f"Unloading diarization model from cache (disabled): {model_name}") 350 | unload_model_object(removed_model) 351 | 352 | return diarize_model -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | --------------------------------------------------------------------------------