Skip to content

CosyVoice

A multilingual speech model with full-stack inference, training and deployment

Visit official site

What CosyVoice is

The model is described as a multilingual large speech generation model with full-stack capabilities from inference to training and deployment. The model is developed by the FunAudioLLM team under the Apache-2.0 license and produces natural, fluent speech close to a human voice.

What you can do with it

  • Generate speech in Chinese, English, Japanese, Cantonese and Korean
  • Run zero-shot, cross-lingual and instruct inference
  • Imitate a specific voice style with SFT voice style conversion
  • Demonstrate the model through a Web interface or deploy it with Docker

Application scenarios

  • Build a multilingual virtual assistant that answers user questions
  • Integrate into education software to voice teaching content for learners of different languages
  • Generate automatic multilingual voice notices or reminders in an enterprise system

Getting started

  • Clone the CosyVoice repository to a local environment
  • Install the required dependencies and environment following the installation guide
  • Download pre-trained models or train a model from scratch with the provided scripts
  • Run speech generation inference tests through example scripts or the Web interface

Pros & cons

✓ What we like

  • Full-stack scripts cover inference, training and deployment
  • A Web interface allows a quick demonstration

! What to watch out for

  • Inference is offered through zero-shot, cross-lingual and instruct modes
  • Using the model starts with cloning the repository and installing dependencies

FAQ

Which languages are named?

The material names Chinese, English, Japanese, Cantonese and Korean among the supported languages.

What does full-stack cover?

The material describes capabilities from inference through training to deployment, with training and inference scripts.

Last reviewed: 2026-11-15

More AI audio tools tools

View all →

How we review