whatmodel.sh / tts / CosyVoice2-0.5B
CosyVoice2-0.5B
↗ share
+ contribute
CosyVoice2-0.5B is an open-weights text-to-speech model from Alibaba (FunAudioLLM), released December 2024. Its published size is 0.5B parameters. Licence: apache-2.0.
Catalogue identity notes
FunAudioLLM/CosyVoice2-0.5B, the only released CosyVoice 2 checkpoint. Distinct from CosyVoice-300M, Fun-CosyVoice3-0.5B and the hosted CosyVoice v3 Plus API (cosyvoice-v3-plus).
Score within the Vapi Humanness · Bradley–Terry rating, 2026-09-25 snapshot comparison set only.
$ how it compares
The text-to-speech models ranked either side of CosyVoice2-0.5B, out of 148 with a score, then more from Alibaba (FunAudioLLM).
| # | model | score | where | ||
|---|---|---|---|---|---|
| 47 |  Eleven Turbo v2 | 61±3 | api $50.00/M chars | ||
| 48 |  Inworld Realtime TTS-2 | 61±2 | api $25.00/M chars | ||
| 49 | CosyVoice2-0.5B | 61±1 | local | ||
| 50 |  E2 TTS Base (F5-TTS reproduction) | 60±2 | local | ||
| 51 |  MiniMax Speech-02-Turbo | 60±3 | api $60.00/M chars | ||
| more from Alibaba (FunAudioLLM) | |||||
| 28 | Fun-CosyVoice3-0.5B-2512 | 64±1 | local | ||
| 68 | CosyVoice-300M | 54±3 | local | ||
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
voice cloning▁·▁▁▁▁▁▁▇█▃▃ #28 of 37 · 34% of its score · 5 tests
Trails Higgs Audio V2, F5-TTS Base (v0), MaskGCT here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Seed-TTS test-en · speaker similarity (WavLM-large SV) +2 | 65.4% ✎ | █·█·██·▄·▄·▄ beats 1/10 | 14% |
| Seed-TTS test-zh · speaker similarity (WavLM-large SV) +1 | 75.0% ✎ | ▃··▃█▅▅··▃·▃ beats 1/10 | 8.9% |
| Seed-TTS test-en · Whisper large-v3 WER +2 | 3.2% ✎ | ▃··▃▅··▅·█▃▅ beats 0/11 | 5.4% |
| Seed-TTS test-zh · Paraformer CER +1 | 1.3% ✎ | ▂··▄··▂▂█·▄▂ beats 1/11 | 2.7% |
| Seed-TTS test-en · human speaker-similarity MOS (SMOS) | 3.07 / 5 ✎ | ▄········█·▄ beats 1/3 | 2.7% |
other tests17 not in a task · 66% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| CV3-Eval hard-en · WER +1 | 13.9% ✎ | █··██·██████ beats 0/8 | 10% |
| CV3-Eval hard-en · speaker similarity (ERes2Net SV) +1 | 70.9% ✎ | ▄█▄·▄··▄··▄█ beats 5/8 | 9.1% |
| CV3-Eval EN (zero-shot English) · WER +1 | 7.6% ✎ | ▃·······▃▅█▅ beats 1/8 | 8.3% |
| CV3-Eval ZH (zero-shot Chinese) · CER | 4.0% ✎ | █▄···█▄·█··▄ beats 1/8 | 5.8% |
| CV3-Eval hard-zh · CER | 8.3% ✎ | ▃▃·▃··▃▃█··▃ beats 6/8 | 4.7% |
| CV3-Eval hard-en · DNSMOS (predicted quality) | 3.95 / 5 ✎ | ▄·····█·▄▄·█ beats 5/6 | 4.1% |
| Raon-OpenTTS-Eval · overall speaker similarity (WavLM) | 60.3% ✎ | ▄·█·█▄·····▄ beats 0/6 | 3.8% |
| Raon-OpenTTS-Eval · human speaker-similarity MOS (overall) | 3.53 / 5 ✎ | █··██······▄ beats 0/6 | 3.5% |
9 more tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| VoxCPM listening test · Chinese speaker-similarity MOS (S-MOS) | 4.01 / 5 ✎ | █·······██·█ beats 1/3 | 3.3% |
| VoxCPM listening test · English speaker-similarity MOS (S-MOS) | 3.97 / 5 ✎ | ██········██ beats 0/3 | 3.2% |
| VoxCPM listening test · English naturalness MOS (N-MOS) | 4.14 / 5 ✎ | █······█··██ beats 3/3 | 3.0% |
| VoxCPM listening test · Chinese naturalness MOS (N-MOS) | 3.38 / 5 ✎ | ██·······█·█ beats 1/3 | 2.6% |
| Raon-OpenTTS-Eval · overall WER (4 acoustic conditions) | 11.0% ✎ | ▄··▄█····▄·█ beats 1/6 | 1.8% |
| RVCBench LibriTTS clean voice cloning · ECAPA speaker similarity | 58.0% ✎ | ▃·····▂▂▂▃█▆ beats 13/15 | 1.6% |
| CodecMOS-Accent VCTK voice clone · human naturalness MOS (S-NAT) | 4.43 / 5 ✎ | █··▄▄██▄·▄▄▄ beats 11/11 | 0.3% |
| CodecMOS-Accent VCTK voice clone · human speaker-similarity MOS (S-SPK-SIM) | 4.24 / 5 ✎ | ▄▄█·▄▄▄▄▄·▄█ beats 7/11 | 0.2% |
| RVCBench LibriTTS clean voice cloning · Whisper-medium WER | 5.0% ✎ | ▂▂▂··▂·▃▃▃█▂ beats 14/15 | 0.0% |
$ where to run it
Other ways to run it
- Run it as published · needs ~4.07 GB · from huggingface.co/FunAudioLLM (official)
git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git && cd CosyVoice && git submodule update --init --recursive && conda create -n cosyvoice -y python=3.10 && conda activate cosyvoice && pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host=mirrors.aliyun.com && python -c "from huggingface_hub import snapshot_download; snapshot_download('FunAudioLLM/CosyVoice2-0.5B', local_dir='pretrained_models/CosyVoice2-0.5B')" && python -c "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'); [torchaudio.save('zero_shot_{}.wav'.format(i), j['tts_speech'], cosyvoice.sample_rate) for i, j in enumerate(cosyvoice.inference_zero_shot('CosyVoice is undergoing a comprehensive upgrade, providing more accurate, stable, faster, and better voice generation capabilities.', '希望你以后能够做的比我还好呦。', './asset/zero_shot_prompt.wav'))]"
Official weights and documentation: huggingface.co/FunAudioLLM ↗
$ how the score was worked out · UltraEval-Audio (OpenBMB) · T… 40%, Raon-OpenTTS: Open Models and… 34%, VoxCPM: Tokenizer-Free TTS fo… 12% and 4 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json