CosyVoice

Why we included this project

CosyVoice is a full-stack multilingual text-to-speech system built on large language models. It does zero-shot voice cloning across languages, so a short reference clip in one language can drive speech in Chinese, English, Japanese, Korean, or several European languages, and it also covers a range of Chinese dialects. The repo includes inference, training, and serving code, so you can move from a working example to a real pipeline without assembling separate pieces yourself. Streaming output with latency as low as 150ms and pronunciation inpainting for Chinese pinyin and English phonemes make it practical for voice agents, audiobook generation, and localized content. For teams that need to ship TTS, the combination of naturalness and production-ready code is rare.