HmZyy/kabyle_transcriber_backend

# Whisper Kabyle Audio Transcription WebSocket Server This project provides a real-time WebSocket (WSS) server for streaming audio transcription using a fine-tuned Whisper model. It supports SSL for secure communication and is designed to transcribe Kabyle (or any supported) language audio sent from WebSocket clients. --- ## 🔧 Features - Real-time transcription over WebSocket - WSS (WebSocket Secure) support using SSL certificates - Fine-tuned Whisper model support (`safetensors` or `pytorch_model.bin`) - JSON-based client communication - Efficient 30s chunk-based audio transcription with 1s overlap - Includes connection state updates and error reporting --- ## 🧪 Quick Setup ### 1. Clone the Repository ```bash git clone github.com cd kabyle_transcriber_backend ``` ### 2. Create Virtual Environment ```bash python3 -m venv venv source venv/bin/activate ``` ### 3. Install Python Dependencies ```bash pip install -r requirements.txt ``` > Ensure you have `ffmpeg` installed for `librosa` audio loading. --- ## 📦 Install `cloudflared` (Optional: Expose localhost to public HTTPS) ```bash curl -L github.com -o cloudflared.deb sudo dpkg -i cloudflared.deb ``` Then run: ```bash cloudflared tunnel --url localhost ``` --- ## 🚀 Run the Server ```bash python3 main.py --host 127.0.0.1 --port 16391 --checkpoint ./checkpoint ``` ## 🧠 WebSocket Message Format ### Audio Message ```json { "type": "audio", "audio_id": "optional-uuid", "format": "base64", "audio_data": " " } ``` ### Ping Message ```json { "type": "ping" } ``` --- ## 📥 Response Format ### Transcription ```json { "type": "transcription", "transcription": "your transcribed text", "timestamp": "ISO 8601 timestamp", "audio_id": "optional-uuid" } ``` ### State Update ```json { "type": "state", "state": "processing | ready | pong | error", "message": "status …