Setup & Installation
Or with the ClawHub CLI, for registry-managed skill folders outside a full OpenClaw workspace:
What This Skill Does
CosyVoice3 runs Alibaba's text-to-speech model locally on macOS Apple Silicon. It supports 9 languages, 18+ Chinese dialects, and zero-shot voice cloning from short audio clips. All inference happens on-device with no network calls required.
Running fully offline on Apple Silicon avoids cloud TTS rate limits, latency, and privacy concerns while still supporting multilingual and dialect output most cloud APIs lack.
When to use it
- Clone a speaker's voice from a 5-second recording
- Generate Cantonese or Sichuan dialect speech offline
- Convert English text using a Chinese reference voice
- Add laughter or breathing cues to synthesized audio
- Batch-produce narration WAV files from a script
Example Workflow
Here's how your AI assistant might use this skill in practice.
User asks: generate a WAV file of this Chinese sentence in a cloned voice from reference.wav
- 1Activate the cosyvoice conda environment
- 2Load the Fun-CosyVoice3-0.5B model via AutoModel
- 3Call inference_zero_shot with the target text and reference.wav
- 4Receive audio chunks from the generator
- 5Save the output to a WAV file using torchaudio
A WAV file containing the synthesized speech matching the reference speaker's voice
Requirements
Accounts, API keys, or tools you or your AI assistant may need to set up while using this skill.