Skip to content
← Models
Open weightsFunAudioLLM (Alibaba)/10 trustMetrics-checked

CosyVoice2

Alibaba's open zero-shot TTS with natural-language instruction control over speaking style.

Visit official page →

About

CosyVoice2 is FunAudioLLM's (Alibaba) second-generation text-to-speech model (December 2024): zero-shot voice cloning, cross-lingual synthesis, and the distinctive feature of instruction-based control — describe the desired speaking style, emotion, or dialect in natural language and the model follows it.

Apache 2.0 licensed with checkpoints on Hugging Face and ModelScope. Strong Chinese/English bilingual performance reflecting its Alibaba origins. The 0.5B variant is lightweight enough for modest GPUs. Community reception has been solid, particularly for Chinese-language TTS where it is among the best open options.

Specs

Context window

N/A (text-to-speech model)

Pricing

Free — open weights (Apache 2.0), self-hosted.

Free quota

Free — download and run the weights yourself (Apache 2.0).

API access

Self-hosted (Python inference code on GitHub; Hugging Face and ModelScope checkpoints).

Weights

Open weights

Released

December 1, 2024

Benchmarks

Provider-claimed (FunAudioLLM paper): state-of-the-art zero-shot TTS quality and speaker similarity — treat as provider-claimed. Independent: widely used in the Chinese open-source TTS community; no standardized third-party benchmark scores.

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

ttsvoice-cloningdubbingmultilingualaudiobooks

Trust score — 7/10

/10 trustMetrics-checked

Scored 7/10 (metrics tier). Apache 2.0 from a serious lab (Alibaba FunAudioLLM), genuinely novel instruction-based style control, and strong bilingual performance. Docked three points: (1) modest direct adoption on Hugging Face (5.9k downloads on the 0.5B checkpoint — popularity concentrates on GitHub/ModelScope); (2) quality claims are paper-reported, not independently benchmarked; (3) documentation is thinner than Western counterparts.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
5.9/10
Maintenance · recency, releases, issue load
5.0/10
Openness · license + docs
8.8/10
Manual review · sanity read + red-flag check
8.0/10
  • · HF downloads 5,855 (peer-normalized)
  • · released 681 days ago
  • · manual review: instruction-based style control is novel; strong bilingual (ZH/EN); popularity concentrates on GitHub/ModelScope over HF; license: Apache-2.0

Sources