PlayVoice/vits_chinese

Combines BERT-derived prosody embeddings with VITS architecture to capture natural pauses from linguistic structure, while incorporating NaturalSpeech's inference loss to reduce audio artifacts. Supports both non-streaming ONNX export and chunked streaming inference via encoder-decoder decomposition, with optional knowledge distillation enabling 3× speedup on a 53M student model. Targets Chinese TTS with multi-speaker capability and includes text preprocessing pipelines for phoneme conversion and numerical text normalization.

1,227 stars. No commits in the last 6 months.