Voce

La voce — 声音

A minimal voice message component for AI chat interfaces, with SenseVoice-powered transcription and emotion detection.

Users speak → browser records → server transcribes (with emotion) → chat renders a voice bar.

Features

• 🎙️ Browser-native recording (MediaRecorder API)
• 🌊 Real-time waveform visualization during recording
• 📝 Server-side transcription via SenseVoice (50+ languages, 15x faster than Whisper)
• 😊 Emotion detection (happy / sad / angry / neutral)
• 🎵 Playable voice bar with waveform progress and duration
• 📋 Transcript toggle ("文" button to view transcribed text)
• 💾 Draft mode: record multiple voice messages before sending

Quick Start

pip install funasr torch torchaudio
python server.py
# Open http://localhost:3456

How It Works

1. Record: Click the mic button → browser uses MediaRecorder to capture audio
2. Upload: Audio blob is uploaded to /upload endpoint
3. Transcribe: Server sends audio to SenseVoice model → returns text + emotion tag
4. Render: Frontend displays a playable voice bar with waveform animation

Tech Stack

• Frontend: Vanilla JS + MediaRecorder API + Web Audio API (for waveform)
• Backend: Python HTTP server + FunASR/SenseVoice
• No frameworks. No build step. Copy and run.

Credits

Built by Cu & Lunedì · cunedi.uk

License

MIT