# CosyVoice

A multilingual large voice generation model with a full pipeline for inference, training and deployment, supporting zero-shot multilingual and cross-lingual voice cloning.

CosyVoice is a multilingual, large-language-model-based text-to-speech system, providing a full-stack pipeline for inference, training and deployment. Its latest version targets zero-shot multilingual speech synthesis, covering nine common languages and more than 18 Chinese dialects and accents, alongside multilingual and cross-lingual zero-shot voice cloning.

It supports pronunciation inpainting for Chinese Pinyin and English CMU phonemes for finer control, text normalization for numbers and special symbols without a separate frontend module, both streaming text input and streaming audio output with latency as low as 150 milliseconds, and instructable generation covering language, dialect, emotion, speed and volume.

It is written in Python and licensed Apache-2.0, with released checkpoints available through ModelScope and Hugging Face.

Website: https://github.com/QwenAudio/CosyVoice