A powerful local text-to-speech engine by Alibaba optimized for macOS Apple Silicon with advanced zero-shot voice cloning and dialect support.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install cosyvoice3-macos
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install cosyvoice3-macos using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is CosyVoice3?
CosyVoice3 is a state-of-the-art text-to-speech (TTS) system based on large language models, specifically optimized for high-performance execution on macOS Apple Silicon. As a key addition to the library of Openclaw Skills, it enables developers to generate natural-sounding speech across 9 languages and over 18 Chinese dialects without relying on external cloud APIs. The skill excels in producing expressive audio with fine-grained control over emotions and prosody.
By integrating this skill into your workflow, you gain access to zero-shot voice cloning, allowing you to replicate any voice from just a few seconds of reference audio. Whether you need cross-lingual synthesis or offline inference, CosyVoice3 provides a robust, privacy-focused solution for high-quality audio generation directly on your local machine using M-series chips.
CosyVoice3 Use Cases
- Generating natural-sounding local TTS for Chinese, English, Japanese, and Korean.
- Cloning specific voices from short 3-10 second reference audio clips for personalized assistants.
- Creating audio content with specific Chinese dialects like Cantonese, Sichuanese, or Shanghainese.
- Implementing offline speech synthesis in applications where data privacy is a top priority.
- Producing expressive narration using text tags for laughter, breathing, and emphasis.
How CosyVoice3 Works
- The user initiates a TTS request through the CLI or integrated Python scripts provided in the skill.
- The system loads the Fun-CosyVoice3-0.5B model into local memory during an initial warmup phase.
- For voice cloning tasks, the engine analyzes a provided reference audio file to extract acoustic characteristics.
- Input text is processed, supporting special markers like <|endofprompt|> and prosody tags.
- The inference engine generates the audio waveform using Apple Silicon's CPU optimization (PyTorch).
- The final output is saved as a high-quality WAV file at the specified local destination.
CosyVoice3 Setup
To get started with this skill from the Openclaw Skills collection, ensure you have an Apple Silicon Mac and Conda installed. Run the following commands:
# Navigate to the skill scripts directory
cd ~/.openclaw/workspace/skills/cosyvoice3/scripts
# Run the automated installation script
bash install.sh
This script handles environment creation, PyTorch installation for macOS, and model downloads (~5GB). To activate the environment manually for usage:
conda activate cosyvoice
CosyVoice3 Data Schema & Taxonomy
CosyVoice3 organizes its models and outputs within a structured repository to ensure high-performance access. The primary data structure includes:
| Component | Path/Location | Description |
|---|---|---|
| Model Files | pretrained_models/ |
Stores the 0.5B and 300M parameter model weights. |
| Reference Assets | asset/ |
Pre-loaded audio for zero-shot and cross-lingual prompts. |
| Skill Scripts | scripts/ |
Contains install.sh and tts.py for automation. |
| Output Files | User-defined | Generated .wav files saved via CLI arguments. |
Reference audio should be 16kHz+ and 3-10 seconds long for the most accurate voice cloning results.
CosyVoice3 Advanced Features
- Zero-Shot Voice Cloning: Replicate any target voice using only a few seconds of source audio without additional training.
- Multi-Dialect Support: Specialized Instruct mode for regional Chinese dialects including Cantonese and Sichuanese.
- Fine-Grained Prosody Control: Insert [laughter], [breath], or tags directly into the text for human-like emotional depth.
- Cross-Lingual Synthesis: Enable a voice from one language (e.g., Chinese) to speak naturally in another (e.g., English).
- Efficient Local Inference: Optimized real-time factor (RTF) of 0.3-0.5 on M1/M2/M3 chips, ensuring fast local production.
SKILL.md
Loading
Related Openclaw Skills
Debt Payoff Plan Comparison
A financial optimization skill that collects debt details to generate and compare various payoff strategies like snowball and avalanche.
Qoder Agent
A high-performance AI coding assistant that executes complex development tasks through the Qoder CLI using a non-interactive Print mode.
Desktop Control for Windows
A comprehensive Windows automation skill for controlling applications, simulating user input, and managing system processes via PowerShell.
Sharesight Skill
A comprehensive tool for managing Sharesight investment portfolios, tracking performance, and automating custom asset data.
AI Trainer & Learning Specialist
An automated knowledge acquisition skill that fetches technical documentation and integrates insights directly into your agent's memory and operational rules.
Chaos Engineer
A specialized AI agent skill for designing controlled failure experiments, managing blast radius, and building antifragile systems.
Featured*
