All models
realtime

CosyVoice 2

Apache-2.0

Alibaba's real-time TTS, 150ms streaming latency

CosyVoice 2 from Alibaba's FunAudioLLM team is engineered for real-time use: streaming synthesis at roughly 150ms first-packet latency with quality nearly identical to offline mode. It supports zero-shot cloning, cross-lingual synthesis, dialect control, and natural-language instructions over emotion and style.

Voices

1 voice
Neutral
Defaultdefault
—
NeutralEmotionalNumbers & Dates

The scripts

Neutral
“The quick brown fox jumps over the lazy dog near the riverbank.”
Emotional
“I can't believe you actually did it. This is incredible!”
Numbers & Dates
“On March 14th, 2025, the team raised $4.2 million at a 38% margin.”

Install

pip install cosyvoice

CosyVoice 2 FAQ

Do I need a GPU to run CosyVoice 2?

For practical speeds, yes — plan on roughly 6 GB of VRAM. If you'd rather not manage a GPU, hosted access to every model in this directory is coming via OpenSpeech Cloud.

Can CosyVoice 2 clone voices?

Yes — CosyVoice 2 supports voice cloning from reference audio.

What license is CosyVoice 2 released under?

CosyVoice 2 is released under the Apache-2.0 license. Always check the repository for the exact terms — some models license code and weights separately.

What languages does CosyVoice 2 support?

CosyVoice 2 supports Chinese, English, Japanese, Korean, Cantonese, WUU, SICHUANESE.

Is there a hosted CosyVoice 2 API?

Yes — OpenSpeech Cloud is launching hosted, pay-per-minute access to CosyVoice 2 and every other model in this directory. No CUDA, Docker, or GPU needed.