CosyVoice
QwenAudio/CosyVoiceReviewed by AI
Checked by RepoPicks Radar on
What is CosyVoice?
A voice generation model from the FunAudioLLM team: it turns text into speech in nine languages and more than eighteen Chinese dialects, copies a voice from a short sample without training, and can follow instructions about emotion, speed or accent. You install Conda and Python, download model files of a few gigabytes from ModelScope or Hugging Face, then run a sample script or a local web page. An NVIDIA card is expected for decent speed, and the faster server setups (vLLM, TensorRT, Docker) are Linux only.
CosyVoice is free and open source (Apache-2.0) and can replace ElevenLabs, Murf, and Fish Audio.
How easy is it to install CosyVoice?
D · terminal, several tools or a specific graphics card · Windows
- Windows D
- macOS D
- Linux D
On every system it is terminal work: clone with submodules, install Conda, create an environment, pip install a long list of packages, download several model folders by script and only then run "python example.py" or "python3 webui.py". Linux is the best supported one: the deepspeed, onnxruntime-gpu, TensorRT and Docker pieces in requirements.txt and in the README only apply there, and sox may need a system package.
Coming soon: install in 1 click with the RepoPicks app.
CosyVoice on GitHub
- Stars
- 24,356on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +64stars in 6 days
Stars over time
Facts
- Runs on
- Windows, macOS, Linux
- You need to install
- Python, NVIDIA graphics card, Other tools
- To build from source
- Python, NVIDIA graphics card
- Disk space
- 1.2 GB–5.1 GB · estimate
- Accounts or keys
- None
- Cost
- Free · Code under Apache-2.0 and model files free to download; the electricity and the graphics card are yours.
- License
- Apache-2.0 (Permissive)
- Can replace
- ElevenLabs, Murf, Fish Audio
- Last change on GitHub
- May 25, 2026
CosyVoice risk signals (6)
- AttentionDependencies with known vulnerabilities
Pinned dependencies with known vulnerabilities (OSV): diffusers: GHSA-7wx4-6vff-v64p, GHSA-98h9-4798-4q5v, PYSEC-2026-2446; gdown: GHSA-76hw-p97h-883f, PYSEC-2026-2158; gradio: GHSA-39mp-8hj3-5c49, GHSA-5cpq-9538-jm2j, GHSA-6655-8ph2-63j3; hydra-core: GHSA-2cp2-2r3c-7p7r, GHSA-c3wx-c55w-pxjq, PYSEC-2026-3850; lightning: GHSA-cgwc-qvrx-rf7f, GHSA-mr7h-w2qc-ffc2, GHSA-qqmf-gpg7-g8gw; modelscope: GHSA-fhhq-h4hg-549x, PYSEC-2026-2663; onnx: GHSA-3r9x-f23j-gc73, GHSA-48m4-24cq-37qf, GHSA-538c-55jv-c5g9; protobuf: GHSA-7gcm-g887-7qv7, GHSA-8qvm-5x2c-j2w7, PYSEC-2026-1805; pyarrow: GHSA-rgxp-2hwp-jwgg, PYSEC-2026-113; torch: GHSA-3749-ghw9-m3mg, GHSA-53q9-r3pm-6pq6, GHSA-63cw-57p8-fm3p; transformers: GHSA-29pf-2h5f-8g72, GHSA-37mw-44qp-f5jm, GHSA-4w7r-h757-3r74
automatic check
- Attentionvoice cloning of real people
The project advertises "multi-lingual/cross-lingual zero-shot voice cloning" from a short sample, and its own disclaimer says "The content provided above is for academic purposes only" and asks people to write in if any example infringes their rights; copying somebody's voice is restricted in several countries.
AI review
- Inforepository is a copy of the original project
This repository is QwenAudio/CosyVoice, but every link in the README points to FunAudioLLM/CosyVoice and the install step says "git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git", so the instructions build the other repository, not this one.
AI review
- Infoinstall from a mirror with certificate check off
README install command: "pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host=mirrors.aliyun.com" — the packages come from an Alibaba mirror and the --trusted-host flag turns off the certificate check for it.
AI review
- Infolarge model downloads from outside github
The README tells you to fetch Fun-CosyVoice3-0.5B, CosyVoice2-0.5B, CosyVoice-300M, -SFT, -Instruct and the ttsfrd resource from ModelScope or Hugging Face, and the optional ttsfrd step installs two .whl files that come inside that download.
AI review
- Infono releases and many open issues
The repository has no GitHub releases, 640 open issues and the last commit is from 2026-05-25 (168 days before the check), so you always install whatever the main branch happens to be.
AI review
Evidence is shown in its original language.
What Radar could not check about CosyVoice
- Whether it runs on Windows or Mac in practice, and how slow it is there without an NVIDIA card.
- Real disk use: the checks estimate 1.2-5.2 GB, and that depends on how many of the model folders you download.
- The quality and speed claims (150 ms latency, the scores in the evaluation table) were measured by the authors.
- Who can reach the web page started by webui.py, which has no login of its own.
- How good the voices are in languages other than Chinese and English.