CosyVoice
github.com
- Free plan
- Linux
- API
- Self-hosted
What CosyVoice is
CosyVoice is a multilingual text-to-speech system for inference, training, and deployment. Fun-CosyVoice 3.0 is designed for zero-shot speech synthesis across nine languages and more than 18 Chinese dialects or accents, and supports multilingual and cross-lingual voice cloning. Users can give instructions for language, dialect, emotion, speed, and volume, and specify pronunciation with Chinese Pinyin or English CMU phonemes. The system can read numbers, special symbols, and varied text formats without a traditional frontend module. Its text-in, audio-out streaming has stated latency as low as 150 ms. The repository documents Conda installation with Python 3.10, pretrained model downloads through ModelScope or Hugging Face, and Docker deployment with gRPC or FastAPI. It also documents NVIDIA Triton and TensorRT-LLM acceleration. The software is licensed under Apache License 2.0, with source and models available for self-managed installation and deployment; no paid plans are listed. The repository describes its content as for academic purposes and technical demonstration.
Who it suits
CosyVoice suits developers and researchers looking to run or deploy multilingual speech synthesis themselves. The repository describes the project as intended for academic purposes and to demonstrate technical capabilities.
Good to know, and what to watch
- Supports nine languages and over 18 Chinese dialects or accents.
- Includes multilingual and cross-lingual voice cloning.
- Streaming latency is stated as low as 150 ms.
- Apache License 2.0 software with self-managed deployment.
- Documented installation uses Python 3.10.
- The optional ttsfrd normalization wheel is specified for Linux x86_64.
Plans
Every plan and priceOpen-source software
Free plan
The verdict
CosyVoice offers controls for multilingual synthesis and several documented deployment paths. It requires a self-managed setup, and the repository frames its content around academic use and technical demonstration.
What the maker publishes
- API access
- Yes source
- API interfaces
- The deployment instructions include FastAPI and gRPC servers and clients. source
- Commercial use
- Yes source
- Deployment options
- The repository documents Docker deployment with gRPC or FastAPI, as well as NVIDIA Triton and TensorRT-LLM acceleration. source
- Export formats
- WAV source
- Free plan
- Yes source
- Installation
- The repository documents installation with Conda and Python 3.10, and offers pretrained model downloads through ModelScope or Hugging Face. source
- Intended use
- The repository says its content is for academic purposes and to demonstrate technical capabilities. source
- Languages and dialects
- Version 3.0 covers nine languages and more than 18 Chinese dialects or accents, with multilingual and cross-lingual zero-shot voice cloning. source
- License
- The repository is licensed under Apache License 2.0. source
- Platforms
- linux,api,self_hosted source
- Pronunciation control
- The system supports pronunciation inpainting with Chinese Pinyin and English CMU phonemes. source
- Purpose
- CosyVoice is a multilingual text-to-speech system that provides inference, training, and deployment capabilities. source
- Runtime requirements
- The documented installation uses Python 3.10; the optional ttsfrd normalization wheel is specified for Linux x86_64. source
- Streaming
- It supports text-in and audio-out streaming, with latency stated as low as 150 ms. source
- Support
- The maintainers direct users to GitHub Issues and an official Dingding chat group for discussion. source
- Text normalization
- It supports reading numbers, special symbols, and varied text formats without a traditional frontend module. source
- vLLM compatibility
- The repository says CosyVoice 2 and 3 support vLLM 0.11.x or newer and vLLM 0.9.0, while versions between those releases are untested. source
- Voice cloning
- Yes source
- Voice instructions
- Users can provide instructions for language, dialect, emotion, speed, and volume. source
- Zero-shot synthesis
- Fun-CosyVoice 3.0 is designed for zero-shot multilingual speech synthesis. source
Checked against the maker's pages on 3 Oct 2026.