CosyVoice 2: Features, Demo, and CosyVoice 2 vs 3 Compared

What is CosyVoice 2?

CosyVoice 2 is a zero-shot, streaming text-to-speech model. The paper describes it as a systematic upgrade of CosyVoice 1 aimed at low latency for voice chat and other interactive apps.

• Built on a pre-trained LLM

  Its text-speech language model is the pre-trained Qwen2.5-0.5B LLM. CosyVoice 1's separate text encoder and the speaker embedding were removed from the language model.

  CosyVoice 2 paper (arXiv 2412.10117)

• FSQ speech tokens at 25 Hz

  Finite scalar quantization replaces vector quantization in the speech tokenizer to use the codebook better. The tokenizer produces 25 speech tokens per second.

  CosyVoice 2 paper (arXiv 2412.10117)

• Streaming and non-streaming in one model

  A chunk-aware causal flow matching model serves both modes. The paper reports virtually lossless quality in streaming mode compared with offline synthesis.

  CosyVoice 2 paper (arXiv 2412.10117)

• Instructions and fine-grained control

  Trained with 1,500 hours of instructed data: natural-language instructions for emotion, speaking rate, role-play and dialect, plus tags such as [laughter], [breath] and <strong> for emphasis.

  CosyVoice 2 paper (arXiv 2412.10117)

• Zero-shot and cross-lingual cloning

  Clones a voice from a reference clip. The repository's example.py shows zero-shot, cross-lingual, instruct and bi-streaming (text fed in as it is generated) calls for CosyVoice2-0.5B.

  QwenAudio/CosyVoice example.py

• Open weights since December 2024

  CosyVoice2-0.5B was released in December 2024 under Apache-2.0. The README roadmap adds vLLM support (May 2025) and a TensorRT-LLM runtime with GRPO training support (August 2025).

  README roadmap

CosyVoice 2 vs CosyVoice 3

How the open CosyVoice 2 release compares with CosyVoice 3 and its open release, Fun-CosyVoice 3.0. Each cell cites its source.

Benchmark numbers are from the evaluation table in the official README, which the CosyVoice team reports for its own models. The CosyVoice 3 paper describes its data growth as "from ten thousand hours to one million hours"; the CosyVoice 2 row uses the CosyVoice 2 paper's own training-data table.

What else changed in the third generation, and how 3.0 differs from the hosted 3.5 models, is in our CosyVoice 3 guide.

Download and run CosyVoice 2

After installing the repository (Python 3.10 Conda environment), download the weights and run the official example. Both snippets are copied from the README and example.py.

# modelscope SDK model download
from modelscope import snapshot_download
snapshot_download('iic/CosyVoice2-0.5B', local_dir='pretrained_models/CosyVoice2-0.5B')

# for oversea users, huggingface SDK model download
from huggingface_hub import snapshot_download
snapshot_download('FunAudioLLM/CosyVoice2-0.5B', local_dir='pretrained_models/CosyVoice2-0.5B')

import sys
sys.path.append('third_party/Matcha-TTS')
from cosyvoice.cli.cosyvoice import AutoModel
import torchaudio

cosyvoice = AutoModel(model_dir='pretrained_models/CosyVoice2-0.5B')

# zero_shot usage
for i, j in enumerate(cosyvoice.inference_zero_shot('收到好友从远方寄来的生日礼物，那份意外的惊喜与深深的祝福让我心中充满了甜蜜的快乐，笑容如花儿般绽放。', '希望你以后能够做的比我还好呦。', './asset/zero_shot_prompt.wav')):
    torchaudio.save('zero_shot_{}.wav'.format(i), j['tts_speech'], cosyvoice.sample_rate)

# instruct usage
for i, j in enumerate(cosyvoice.inference_instruct2('收到好友从远方寄来的生日礼物，那份意外的惊喜与深深的祝福让我心中充满了甜蜜的快乐，笑容如花儿般绽放。', '用四川话说这句话<|endofprompt|>', './asset/zero_shot_prompt.wav')):
    torchaudio.save('instruct_{}.wav'.format(i), j['tts_speech'], cosyvoice.sample_rate)

Weights:Hugging Face: CosyVoice2-0.5BModelScope: CosyVoice2-0.5B

Full install steps, including Windows and Docker notes, are in our CosyVoice install guide.

CosyVoice 2 FAQ