CosyVoice

A scalable multilingual zero-shot text-to-speech synthesizer based on large language models, designed for speech generation in the wild.

Fun-CosyVoice 3.0 (Latest) CosyVoice 2.0 CosyVoice 1.0

Key Features

State-of-the-art TTS with powerful capabilities

🌐

Multi-Language Support

Covers 9 languages (Chinese, English, Japanese, Korean, German, Spanish, French, Italian, Russian) and 18+ Chinese dialects including Cantonese, Minnan, Sichuan, Shanghai, and more.

🎤

Zero-Shot Voice Cloning

Clone any voice with just a short audio prompt. Supports both multi-lingual and cross-lingual zero-shot synthesis with high speaker similarity.

Bi-Streaming

Supports both text-in streaming and audio-out streaming, achieving latency as low as 150ms while maintaining high-quality audio output.

🎫

Instruction Control

Supports various instructions for controlling languages, dialects, emotions, speaking speed, volume, and more for fine-grained speech generation.

📚

Text Normalization

Built-in reading of numbers, special symbols, and various text formats without requiring a traditional frontend module.

🔧

Pronunciation Inpainting

Supports Chinese Pinyin and English CMU phoneme inpainting for precise pronunciation control, suitable for production use.

Audio Demos

Listen to samples generated by each version of CosyVoice

Quick Start

Get CosyVoice running in minutes

1 Clone the repository
git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git cd CosyVoice git submodule update --init --recursive
2 Create environment & install dependencies
conda create -n cosyvoice -y python=3.10 conda activate cosyvoice pip install -r requirements.txt
3 Download pretrained models
# Using ModelScope SDK from modelscope import snapshot_download snapshot_download('FunAudioLLM/Fun-CosyVoice3-0.5B-2512', local_dir='pretrained_models/Fun-CosyVoice3-0.5B') # Or using HuggingFace SDK from huggingface_hub import snapshot_download snapshot_download('FunAudioLLM/Fun-CosyVoice3-0.5B-2512', local_dir='pretrained_models/Fun-CosyVoice3-0.5B')
4 Run inference
# Run the example script python example.py # Or start the web demo python3 webui.py --port 50000 --model_dir pretrained_models/CosyVoice-300M

Evaluation

Benchmark results on CV3-Eval test sets

Model Open Source Size test-zh CER%↓ test-zh SS%↑ test-en WER%↓ test-en SS%↑ test-hard CER%↓ test-hard SS%↑
Human-- 1.2675.52.1473.4--
Seed-TTS- 1.1279.62.2576.27.5977.6
F5-TTS0.3B 1.5274.12.0064.78.6771.3
CosyVoice20.5B 1.4575.72.5765.96.8372.4
Index-TTS21.5B 1.0376.52.2370.67.1275.5
VoxCPM0.5B 0.9377.21.8572.98.8773.0
Fun-CosyVoice3-0.5B0.5B 1.2178.02.2471.86.7175.8
Fun-CosyVoice3-0.5B RL0.5B 0.8177.41.6869.55.4475.0

Available Models

Download from ModelScope or HuggingFace

Fun-CosyVoice3-0.5B

Latest and recommended model with best overall performance. 0.5B parameters.

ModelScope · HuggingFace

CosyVoice2-0.5B

Streaming TTS with 25Hz codec. Supports real-time speech synthesis.

ModelScope · HuggingFace

CosyVoice-300M

Original model with SFT and Instruct variants available.

ModelScope · HuggingFace

FunAudioLLM Ecosystem

CosyVoice is part of a complete speech AI toolkit

Publications

Read the research behind CosyVoice