whisper-api-server/README.md

Whisper Speech-to-Text API Service

Overview

This project is a local API server compatible with OpenAI's API for transcribing audio to text using the Whisper model. It's designed to run as a system service, loading the Whisper model into memory at startup and handling transcription requests via REST API.

Features

• Audio Transcription: Supports various input methods:
  • Local server files
  • Files accessible via URL
  • Base64-encoded files
  • Multipart form data
• OpenAI API Compatibility: Works with /v1/audio/transcriptions and /v1/models endpoints.
• Audio Preprocessing: Converts audio to WAV, normalizes, and adds silence.
• Hardware Support: Utilizes GPU (CUDA, MPS) or CPU.
• Logging: Tracks all operations.
• Health Check: Includes a health check endpoint.

Quick Start

1. Edit the Configuration File (config.json):

   • Set the path to your Whisper model (model_path).
   • Configure other parameters like language (language), chunk size (chunk_length_s), batch size (batch_size), and audio normalization settings.

   {
       "service_port": 5042,
       "model_path": "/path/to/your/whisper-model",
       "language": "english",
       "chunk_length_s": 30,
       "batch_size": 16,
       "max_new_tokens": 256,
       "return_timestamps": false,
       "norm_level": "-0.5",
       "compand_params": "0.3,1 -90,-90,-70,-70,-60,-20,0,0 -5 0 0.2"
   }
   

2. Run the Server:

   • Simply execute the server.sh script:

   ./server.sh
   

   • If you need to update the environment, use:

   ./server.sh --update
   

3. Use the API:

   • Once the server is running, you can send transcription requests.
   • Example request (curl):

   curl -X POST -F file=@audio.wav http://localhost:5042/v1/audio/transcriptions
   

Enjoy seamless audio-to-text transcription with your local Whisper API server!