Meowl Voice

Voice cloning running locally on Apple Silicon. Record ~30 seconds of speech, then generate anything in that voice.

Built with Qwen3-TTS and MLX. FastAPI backend, React frontend.

Requirements

• macOS with Apple Silicon (M1+)
• Python 3.10+ and uv
• Node.js 18+
• ffmpeg (brew install ffmpeg)

Setup

# Install dependencies
make install

# Run (builds frontend, serves everything on :8000)
make run

Then open http://localhost:8000.

For development (hot-reload frontend on :5173, API on :8000):

make dev

How it works

1. Record — Read the script out loud (~10-30 seconds). The app captures your voice via the browser mic.
2. Train — The recording is processed into a voice profile using Qwen3-TTS speaker embeddings.
3. Generate — Type anything and generate speech in the cloned voice.

The model runs entirely on-device via MLX on the Apple GPU. No data leaves your machine.

Models

Uses cr2k2/Qwen3-TTS-12Hz-1.7B-Base-fp32 by default (~4.5GB, downloaded automatically on first run). The 0.6B model is available but not recommended for voice cloning quality.

Performance

On Apple Silicon (M5 MacBook Pro), expect ~29 seconds wall time for ~10 seconds of audio (RTF ~2.9x). The first generation is slower due to MLX graph compilation — a warmup pass runs automatically at startup.

Configuration

Environment variables:

Variable	Default	Description
MEOWLVOICE_BACKEND	mlx	Backend engine (mlx or pytorch)
MEOWLVOICE_API_HOST	127.0.0.1	API bind address
MEOWLVOICE_API_PORT	8000	API port
MEOWLVOICE_CORS_ORIGINS	localhost:5173,...	Allowed CORS origins
MEOWLVOICE_LOG_LEVEL	INFO	Log level

Tech stack

• Backend: FastAPI + MLX Audio + ffmpeg
• Frontend: React + Vite
• Model: Qwen3-TTS 1.7B (fp32 MLX conversion)

License

MIT