/Whisper-API-Server

Whisper OpenAI Compatible API server

Primary LanguagePythonMIT LicenseMIT

Whisper API Server (OpenAI compatible API)

This is a Flask-based API server that provides speech-to-text transcription functionality using the Faster Whisper library. It supports multiple pre-trained models and allows users to transcribe audio files in various formats.

This server's API is fully compatible original OpenAI's whisper API spec.

Author

  • code by Jioh L. Jung
  • Code and manual readme file was created with Claude3 opus(AI assisted).

License

  • MIT License

Requirements

  • Python 3.6 or higher
  • Flask
  • faster-whisper
  • soundfile
  • torch

Installation

  • Clone the repository or download the source code.
  • Install the required dependencies using pip:
pip install -r requirements.txt
  • Download the pre-trained models you want to use and place them in the appropriate directory.

Usage

Start the API server by running the following command:

python api-server.py

  • The server will start running on http://0.0.0.0:2638.

  • Send a POST request to the /v1/audio/transcriptions endpoint with the following parameters:

  • file (required): The audio file to be transcribed. Supported formats: WAV, MP3, MP4, MPEG, MPGA, M4A, WEBM.

  • model (optional): The name of the pre-trained model to use for transcription. Default: tiny.en. Available models are defined in the MODELS_LIST variable.

  • min_silence_duration_ms (optional): The minimum duration of silence (in milliseconds) to be considered as a separate segment. Default: 500.

  • temperature (optional): The temperature value for the transcription model. Default: 0.0.

  • language (optional): The language of the audio file. Default: auto (automatic detection).

  • response_format (optional): The format of the response. Supported formats: json, text, srt, verbose_json, vtt. Default: json.

  • The API will respond with the transcription result in the specified format.

Transcription Process

  • The API server receives the audio file and reads it into memory using io.BytesIO().
  • The audio data is then converted to WAV format using the soundfile library.
  • The Faster Whisper model is loaded based on the specified model parameter. The server attempts to use the available device in the following order: CUDA, MPS (for macOS), CPU.
  • The audio data is transcribed using the loaded model with the specified parameters (min_silence_duration_ms, temperature, language).
  • The transcription result is formatted according to the response_format parameter and returned as the API response.

Supported Response Formats

  • json: Returns the transcription result as a JSON object with a single text field containing the transcribed text.
  • text: Returns the transcribed text as plain text.
  • srt: Returns the transcription result in the SubRip (SRT) subtitle format.
  • verbose_json: Returns a detailed JSON object containing additional information about the transcription, such as segment timestamps, tokens, and probabilities.
  • vtt: Returns the transcription result in the Web Video Text Tracks (WebVTT) format.

Error Handling

The API server handles the following error scenarios:

  • If no file is uploaded, it returns a 400 Bad Request response with an error message.
  • If an unknown model name is specified, it returns a 400 Bad Request response with an error message and the list of available models.
  • If no speech is detected in the audio file, it returns a 400 Bad Request response with an error message. Notes
  • The server runs on http://0.0.0.0:2638 by default. You can modify the host and port parameters in the app.run() function call to change the server's address and port.
  • The available pre-trained models are defined in the MODELS_LIST variable. You can uncomment or add more models as needed.
  • The server attempts to use the available device for model loading in the following order: CUDA, MPS (for macOS), CPU. If a device is not available or fails, it falls back to the next available device.