cosyvoice3

What This Skill Does

CosyVoice3 is a state-of-the-art text-to-speech (TTS) engine by Alibaba, optimized specifically for Apple Silicon (M1/M2/M3) hardware via OpenClaw. This skill provides high-fidelity, natural-sounding speech synthesis supporting 9 major languages (including English, Chinese, Japanese, and Korean) and over 18 distinct Chinese dialects. Unlike cloud-based TTS solutions, this runs entirely locally on your machine, ensuring data privacy and offline accessibility. It features powerful zero-shot voice cloning capabilities, allowing you to synthesize speech that mimics a specific person's timbre using only 3-10 seconds of reference audio. Additionally, it supports cross-lingual synthesis, meaning you can generate English speech using a Chinese voice profile, and provides fine-grained control over prosody, speed, and emotional inflection via text-based tags.

Installation

To install this skill, execute the following command in your OpenClaw environment: clawhub install openclaw/skills/skills/lhuaizhong/cosyvoice3-macos After installation, navigate to /Users/lhz/.openclaw/workspace/skills/cosyvoice3/scripts and run bash install.sh. This process will automatically set up a dedicated Conda environment, configure the necessary PyTorch dependencies for your Apple Silicon hardware, and download the Fun-CosyVoice3-0.5B model weights.

Use Cases

• Professional Voice Overs: Generate high-quality narration for videos or presentations without expensive studio equipment.
• Content Localization: Easily translate and synthesize scripts into multiple languages while maintaining a consistent voice identity.
• Accessibility & Assistive Tech: Create natural, human-like voice feedback for applications or reading assistants.
• Creative AI Projects: Clone voices for character narration in games, animations, or personalized audiobooks.

Example Prompts

1. "Use CosyVoice3 to narrate this article in a calm, professional tone using the default female voice."
2. "Clone my voice from 'reference.wav' and read the following text: 'Hello, this is a test of my synthetic twin.'"
3. "Synthesize this Chinese script into English using the voice from my saved assets, and set the speed to 1.2x."

Tips & Limitations

• Reference Audio: When performing zero-shot cloning, ensure the audio is clear, free of background noise, and 3-10 seconds long for best results.
• Tagging: Always include the <|endofprompt|> token in your reference text segments to help the model distinguish between prompt content and generated output.
• Performance: While Apple Silicon is efficient, generating very long audio clips may take time. Break large blocks of text into smaller paragraphs for faster synthesis.
• Storage: Ensure you have at least 5GB of free space before beginning the installation.