FastWhisperAPI is a web service built with the FastAPI framework, specifically tailored for the accurate and efficient transcription of audio files using the Faster Whisper library. This project is an open-source initiative that leverages the remarkable Faster Whisper model. It is four times faster than openai/whisper while maintaining the same level of accuracy and consuming less memory, whether running on CPU or GPU. The API is built to provide compatibility with the OpenAI API standard, facilitating seamless integration into existing applications that use OpenAI - Whisper.
If you find FastWhisperAPI useful, please consider leaving a star ⭐ or donate.
- FastWhisperAPI is fully compatible with the OpenAI API standard.
- Transcribe audio files asynchronously using a ThreadPoolExecutor.
- Support for submitting multiple files per request.
- Support for multiple languages and model sizes.
- Customizable initial prompt to guide the model's transcription process.
- Voice activity detection filter.
- Customizable response format and timestamp granularities.
- Python 3.8 or greater
- Refer to the Faster Whisper documentation for the GPU requirements here.
-
Clone the repository:
git clone https://github.com/3choff/FastWhisperAPI.git
-
Navigate to the project directory:
cd FastWhisperAPI -
Create a new environment:
python3 -m venv FastWhisperAPI
-
Activate the virtual environment:
-
On Unix/Linux/macOS:
source FastWhisperAPI/bin/activate -
On Windows:
FastWhisperAPI\Scripts\activate
-
-
Install the required packages:
pip install -r requirements.txt
To run the FastAPI app, use the following command:
fastapi run main.pyIf you want to specify a different port, use the --port option followed by the desired port number:
fastapi run main.py --port 5000The API automatically detects the availability of a GPU and configures the device accordingly, either on CPU or CUDA.
If you wish to explicitly run the application on CPU, even if CUDA cores are available, set the FORCE_CPU environment variable to "true":
FORCE_CPU=true fastapi run main.pyThe environment variable is unnecessary if only the CPU is available.
The application will begin running at http://localhost:8000 if the port was not specified, or at http://localhost:PORT_NUMBER if a different port was specified.
To authenticate API requests, set the API key to "dummy_api_key" in your environment.
This API can be dockerized for deployment, and a Dockerfile is included in the repository. Please note that you may need to edit the Dockerfile based on your specific setup and CUDA version installed.
Use the following commands to build and run the container:
Build a Docker container:
docker build -t fastwhisperapi .Run the container
docker run -p 8000:8000 fastwhisperapiAdditionally, it is possible to run the API from a Google Colab environment using the Ngrok service. The Ngrok service will generate a random public URL that you can use to replace localhost in your project. For example, if the URL assigned is https://a3bd-34-171-99-34.ngrok-free.app, the transcription endpoint will be https://a3bd-34-171-99-34.ngrok-free.app/v1/transcriptions.
The Jupyter Notebook for running the API in Colab is also included in the repository.
file: A list of audio files to transcribe. This is a required parameter.model: The size of the model to use for transcription. This is an optional parameter. The options are 'large', 'medium', 'small', 'base', 'tiny'. Default is 'base'.language: This parameter specifies the language of the audio files. It is optional, with accepted values being lowercase ISO-639-1 format (e.g., 'en' for English). If not provided, the system will automatically detect the language.initial_prompt: This optional parameter provides an initial prompt to guide the model's transcription process. It can be used to pass a dictionary of the correct spellings of words and to provide context for better understanding speech, thus maintaining a consistent writing style.vad_filter: Whether to apply a voice activity detection filter. This is an optional parameter. Default is False.min_silence_duration_ms: The minimum duration of silence to be considered as a pause. This is an optional parameter. Default is 1000.response_format: The format of the response. This is an optional parameter. The options are 'text', 'verbose_json'. Default is 'text'.timestamp_granularities: The granularity of the timestamps. This is an optional parameter. The options are 'segment', 'word'. Default is 'segment'. This is a string and not an array like the OpenAI API, and the timestamps will be returned only if the response_format is set to verbose_json.
You can use the following curl command to send a POST request to the /v1/transcriptions endpoint:
curl -X POST "http://localhost:8000/v1/transcriptions" \
-H "accept: application/json" \
-H "Content-Type: multipart/form-data" \
-F "[email protected];type=audio/wav" \
-F "[email protected];type=audio/wav" \
-F "model=base" \
-F "language=en" \
-F "initial_prompt=RoBERTa, Mixtral, Claude 3, Command R+, LLama 3." \
-F "vad_filter=False" \
-F "min_silence_duration_ms=1000" \
-F "response_format=text" \
-F "timestamp_granularities=segment"/: Redirects to the/docsendpoint, which provides a Swagger UI for interactive exploration of the API. You can call and test the API directly from your browser./info: Provides information about the device used for transcription and the parameters./v1/transcriptions: API designed to transcribe audio files.
This project was made possible thanks to:
- Faster Whisper: For providing the transcription model used in this project.
- FastAPI: For the web framework used to build the API.
- AI Anytime: For inspiring this project.
If you find this project helpful and would like to support its development, there are several ways you can contribute:
- Support: Consider donate to support my work.
- Contribute: If you're a developer, feel free to contribute to the project by submitting pull requests or opening issues.
- Spread the Word: Share this project with others who might find it useful.
Your support means a lot and helps keep this project going. Thank you for your contribution!
This project is licensed under the Apache License 2.0.
