whatmodel.sh / tts / CosyVoice-300M

CosyVoice-300M

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

CosyVoice-300M is an open-weights text-to-speech model from Alibaba (FunAudioLLM), released July 2024. Its published size is 0.3B parameters. Licence: apache-2.0.

Catalogue identity notes

FunAudioLLM/CosyVoice-300M (mirrored as model-scope/CosyVoice-300M), the zero-shot base checkpoint; not CosyVoice-300M-SFT/-Instruct or CosyVoice 2/3.

54.4±2.6#55 of 118 text-to-speech modelsfrom 10 benchmark results across 3 sources

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 CosyVoice-300M, out of 148 with a score, then more from Alibaba (FunAudioLLM).

#modelscorewhereRAMreleased
66Higgs Audio V2 Boson AI55±3local~7GB2025-07
67Fasih-TTS-V1 NightPrince (Yahya Elnawasany)55±10local~1.9GB2026-07
68CosyVoice-300M Alibaba (FunAudioLLM)54±3local~2.3GB2024-07
69Chatterbox Multilingual V3 Resemble AI54±10local~3.2GB2026-04
70VibeVoice 1.5B Microsoft53±6local~5.4GB2025-08
more from Alibaba (FunAudioLLM)
28Fun-CosyVoice3-0.5B-2512 Alibaba (FunAudioLLM)64±1local~5.4GB2025-12
49CosyVoice2-0.5B Alibaba (FunAudioLLM)61±1local~4.1GB2024-12

see all text-to-speech models · find the ones that fit your machine

$ what it's good at

voice cloning▁·▁▁▁▁▃▁▇█▃▃ #30 of 37 · 54% of its score · 4 tests

Trails Spark-TTS 0.5B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Seed-TTS test-en · speaker similarity (WavLM-large SV)64.0% ✎▁·▂▂▃█▆▃▅··▂ beats 5/30medium · avg 53%▇▇▇▇▇▇▇▇ 19%at its level
Seed-TTS test-zh · speaker similarity (WavLM-large SV)75.0% ✎▁·▂··▂·█▄·▃▄ beats 5/25medium · avg 62%▇▇▇▇▇▇▇ 17%at its level
Seed-TTS test-en · Whisper large-v3 WER +13.4% ✎▁···▂▃▂▂█▆▄▂ beats 3/35avg 3.6%▇▇▇▇▇ 11%blunt · several runs
Seed-TTS test-zh · Paraformer CER +13.1% ✎▁·▃▁▁▁▂▃▅█·▂ beats 1/30avg 2.1%▇▇▇ 7.0%blunt · several runs
other tests4 not in a task · 43% of its score
testresultvs the others in its rundifficultycountswhy
LibriSpeech-PC test-clean (F5-TTS subset) · speaker similarity (WavLM SV)66.0% ✎▄·········█▄ beats 2/3medium · avg 52%▇▇▇▇▇▇▇▇▇ 21%at its level
VERSA TTS demo · LibriSpeech test-clean UTMOS4.15 / 5 ✎▄·····▄·▄█·▄ beats 5/5easy · avg 4.11 / 5▇▇▇▇▇ 12%typical
LibriSpeech-PC test-clean (F5-TTS 4-10 s subset) · Whisper large-v3 WER3.6% ✎█·····█·█··█ beats 0/3avg 4.1%▇▇▇▇ 9.0%typical
VERSA TTS demo · LibriSpeech test-clean Whisper WER5.0% ✎▄··········█ beats 2/5avg 6.4%▇ 0.5%blunt · noisy

$ where to run it

Other ways to run it

  1. Run it as published · needs ~2.3 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/CosyVoice-300M', local_dir='pretrained_models/CosyVoice-300M')" && 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/CosyVoice-300M'); [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 · F5-TTS: A Fairytaler that Fak… 78%, VERSA toolkit paper: TTS demo… 12%, Amphion MaskGCT README · zero… 7%

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

sourceresultsshare of the scorelicence
F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching (arXiv:2410.06885v1)6▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 78%CC BY 4.0
VERSA toolkit paper: TTS demonstration (arXiv:2412.17667v2)2▇▇▇▇ 12%CC BY 4.0
Amphion MaskGCT README · zero-shot TTS results table2▇▇ 7%MIT
prior · pull toward the middle▇ 3%
weighted average= 54.4

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

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