博客

CosyVoice 3 scales training data to 1M hours and 1.5B parameters, adds a new supervised speech tokenizer and reinforcement-learning post-training, and supports 9 languages plus 18 Chinese dialects. Here is everything that changed.

A side-by-side comparison of CosyVoice 1, 2, and 3 — training data, model size, languages, tokenizer, latency, and voice cloning quality — to help you choose the right version.

A practical guide to installing CosyVoice from GitHub, downloading the pretrained models, and generating your first zero-shot voice clone in Python — plus Docker and web UI options.

A clear walkthrough of how CosyVoice 3 works — the supervised multi-task speech tokenizer, the LLM with chunk-aware flow matching, and the differentiable reward model used for post-training.

How CosyVoice 3 is evaluated — content consistency (WER/CER) and speaker similarity (SS) — what the metrics mean, how it compares to CosyVoice 2, and how to read TTS benchmark numbers.