mstar.model.cosyvoice3.components

mstar.model.cosyvoice3.components#

Modules

dit

CosyVoice3's flow-matching velocity estimator: an F5-TTS style DiT.

onnx_weights

Weights of the ONNX-only front-end models, loaded into PyTorch modules.

sampling

Repetition-aware sampling (RAS) for the CosyVoice3 speech-token LM.

text

CosyVoice3's text tokenizer: the Qwen2 BPE of CosyVoice-BlankEN plus the special tokens CosyVoice3 adds (prompt delimiters, paralinguistic tags, and English and pinyin phonemes for pronunciation fixes such as [j][ǐ]).