myshell-ai/OpenVoice

Separates tone color cloning from fine-grained style control (emotion, accent, rhythm, intonation), enabling zero-shot cross-lingual synthesis without requiring both source and target languages in training data. Built on VITS architecture with a two-stage training strategy that decouples speaker characteristics from linguistic content. V2 adds native support for six languages (English, Spanish, French, Chinese, Japanese, Korean) and improves audio quality through revised training methodology.

36,111 stars. No commits in the last 6 months.