CosyVoice

CosyVoice is the text-to-speech half of Alibaba's FunAudioLLM family, sitting opposite FunASR on the recognition side. The current open-weights checkpoint is Fun-CosyVoice3-0.5B-2512, released in December 2025 at 0.5B parameters under Apache 2.0. That licence covers the weights as well as the training and inference code, which is the single most important fact about the project and the one most English coverage gets wrong. Nine languages are supported (Chinese, English, Japanese, Korean, German, Spanish, French, Italian and Russian) along with more than 18 Chinese dialects and regional accents.

Reported numbers on the model card are 1.21% CER with 78.0 speaker similarity for Chinese and 2.24% WER with 71.8 for English, and the reinforcement-learning post-trained variant brings Chinese CER down to 0.81%. Architecturally it pairs supervised semantic tokens with conditional flow matching for the acoustic stage, and the bi-streaming mode holds first-audio latency around 150ms, fast enough to put a voice on a conversational agent without an audible gap before it speaks. A Triton TensorRT-LLM runtime exists for production serving and CosyVoice2-0.5B has a vLLM path for batch throughput. Zero-shot cloning works from a short reference clip and crosses languages, so a Mandarin reference speaker can be made to read English at usable quality.

Two practical notes for anyone working from either side of the firewall. Weights are mirrored on Hugging Face and ModelScope under the FunAudioLLM organisation, so the pull works wherever you are, though ModelScope is substantially faster from inside the mainland. The repository has also moved between GitHub organisations more than once, which means pinned dependency URLs and a good share of the tutorial write-ups now resolve to nothing. Past installation, the English documentation thins out fast. Fine-tuning recipes, dialect conditioning and speaker embedding questions are answered in Mandarin issue threads and on ModelScope discussion pages, and machine translation handles them badly because the terminology is domain-specific rather than general.