CosyVoice

CosyVoice is the text-to-speech model of Alibaba's FunAudioLLM family, introduced in the July 2024 arXiv paper (2407.05407). It is a scalable multilingual zero-shot TTS synthesizer built on supervised semantic tokens: it generates natural speech from text, clones voices from short reference clips, performs cross-lingual synthesis (a voice sample in one language speaking another), and supports voice conversion and streaming inference.

The models and training/inference/fine-tuning code were open-sourced on Hugging Face, ModelScope, and GitHub, and the line has since grown into CosyVoice 2 (streaming, December 2024) and CosyVoice 3 (2025). Note on dating: the release date here is era-level from the July 2024 paper; verify the exact license on the model card before commercial use.