Back to Gap Map Model components / Speech & audio

CosyVoice

Alibaba Cloud
open weights / Overall score: 3.0

Alibaba's multilingual voice generation line (CosyVoice, CosyVoice 2 and Fun-CosyVoice 3), with zero-shot voice cloning and instruction control over dialect, emotion and speed in 9 languages and more than 18 Chinese dialects. The repository ships inference, training and deployment code.

The repository moved from FunAudioLLM/CosyVoice to QwenAudio/CosyVoice; both names resolve to the same repository. Code and weights are Apache-2.0.

Openness

3 medium confidence
3.0
weights
open (checkpoints on the Hub, ungated)
data
closed (no training corpus named in the card or README)
code
open (inference, training and deployment code in the repository)
license
Apache-2.0 (OSI)

CosyVoice's code and weights are Apache-2.0, OSI-approved, and the repository includes training code. Alibaba does not publish or name the data the checkpoints were trained on, which keeps it at open weights.

Adoption

3 high confidence
3.0

Hugging Face downloads of the two current checkpoints; the older CosyVoice 1 repositories add little.

Capability

3 low confidence
3.0

A leading open Chinese voice-cloning line with dialect coverage few peers match. It is not on the Artificial Analysis TTS leaderboard, so it is placed on its own reported results on a shared public test set, level with Kokoro, rather than on a blind listening test.

Verified 2026-09-27