CosyVoice
Multilingual LLM-based text-to-speech with zero-shot voice cloning, Chinese dialects and streaming output.
Quick answer
CosyVoice is a multilingual LLM-based text-to-speech with zero-shot voice cloning, Chinese dialects and streaming output, listed under Speech & translation. A typical small-business use: generate Chinese or multilingual narration for training videos, product demos and ads. The project publishes no official RAM figure, so we do not estimate a monthly server cost. License: Apache-2.0 (code; the Fun-CosyVoice3 and CosyVoice2 model cards also list Apache-2.0), which allows internal business use.
Last updated · facts verified against official docs on · sources
Activity data comes from GitHub and was last refreshed Nov 2, 2026.
What is CosyVoice?
CosyVoice is the open text-to-speech project behind the Fun-CosyVoice 3.0, CosyVoice 2.0 and 1.0 models. It does zero-shot voice cloning from a short reference clip, covers 9 languages (Chinese, English, Japanese, Korean, German, Spanish, French, Italian, Russian) plus 18+ Chinese dialects and accents, accepts instructions for emotion, speed and dialect, and streams audio with latency as low as 150 ms. It ships training and inference code, a Gradio web demo, FastAPI and gRPC servers, and optional vLLM or TensorRT-LLM acceleration.
Where it fits in a 10–200 person business
- Generate Chinese or multilingual narration for training videos, product demos and ads
- Give a phone bot or in-store kiosk a natural voice, including regional Chinese dialects
- Produce consistent brand voice-overs from a consenting speaker's reference clip
These use cases are our editorial suggestions, not claims made by the project.
Can a business use CosyVoice? (license)
| License | Apache-2.0 (code; the Fun-CosyVoice3 and CosyVoice2 model cards also list Apache-2.0) |
|---|---|
| Type | Permissive: MIT / Apache-2.0: use, modify and sell with few obligations. |
| Internal business use | Allowed under the license text |
| What to watch | Free for internal and commercial use, modification and redistribution. Keep the license and notices. Check each model's card for its license before commercial use. |
This is a plain-English summary of the license text, not legal advice. Third-party models you run with it have their own licenses.
Data & compliance: does CosyVoice phone home?
| Telemetry (usage data sent to the vendor) | Not documented |
|---|---|
| Other calls home by default (updates, marketplaces, default cloud services) | Not documented |
| Single sign-on (OIDC / SAML / LDAP) | Not documented |
| Audit log | Not documented |
| Role-based access control (RBAC) | Not documented |
| Data residency / GDPR notes from the project | Not documented |
Checked against the project's own docs and repository on . "Not documented" means we could not find the answer in official sources; it does not mean the feature is missing or that no data is sent. Third-party services you connect (model APIs, vector databases, search engines) have their own data policies. Compare all apps.
What hardware does CosyVoice need?
The project does not publish a minimum CPU/RAM figure, so we leave the official value blank rather than guess.
GPU: Not stated in the README; the documented Docker server commands run with --runtime=nvidia.
No CPU, RAM, GPU memory or disk figures are published. Install uses a Conda environment with Python 3.10; the optional ttsfrd text-normalisation wheel is Linux x86_64 / Python 3.10 only.
How much does it cost to host CosyVoice per month?
The project publishes no RAM figure we could verify, so we do not estimate a cost.
How to deploy CosyVoice
Method: Docker (self-built image, FastAPI server)
git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git
cd CosyVoice/runtime/python
docker build -t cosyvoice:v1.0 .
docker run -d --runtime=nvidia -p 50000:50000 cosyvoice:v1.0 /bin/bash -c "cd /opt/CosyVoice/CosyVoice/runtime/python/fastapi && python3 server.py --port 50000 --model_dir iic/CosyVoice-300M && sleep infinity"
# web demo instead: python3 webui.py --port 50000 --model_dir pretrained_models/CosyVoice-300MFrom the official instructions: github.com/QwenAudio/CosyVoice#install. Always check the source for the current version before you run it.
Known gotchas when self-hosting CosyVoice
- There is no prebuilt image: you build the Docker image from runtime/python, and the documented server commands use the older CosyVoice-300M model, while the README recommends Fun-CosyVoice3-0.5B for best quality. Change --model_dir accordingly. source
- Models are downloaded separately from ModelScope or Hugging Face into pretrained_models/ before use. source
- vLLM acceleration works only with vLLM 0.9.0 or 0.11.x+ (0.10.x is untested, older is unsupported); the README suggests a separate Conda env because vLLM has strict dependency requirements. source
- The repository moved from FunAudioLLM/CosyVoice to QwenAudio/CosyVoice; old links redirect. source
Alternatives in Speech & translation
- whisper.cpp Fast C/C++ port of OpenAI Whisper speech recognition with a small HTTP transcription server.
- PDFMathTranslate PDF translator that keeps the original layout, formulas, charts and tables.
- Fish Speech Multilingual text-to-speech and voice cloning server (Fish Audio Drama 3 and S2-Pro) with an HTTP API and web UI.
- Meetily Desktop meeting recorder that transcribes and summarizes meetings locally.
- LibreTranslate Self-hosted machine translation API with no dependency on Google or DeepL.
- Anarlog Local-first AI meeting notepad that records device audio without a bot joining the call.
- Speaches OpenAI-compatible speech-to-text and text-to-speech server (faster-whisper, Kokoro, Piper).
- Whisper ASR Webservice Whisper speech-to-text as a REST API in one Docker container.
- Scriberr Self-hosted web app that transcribes audio offline, with speaker labels and AI summaries.
CosyVoice FAQ
How much RAM does CosyVoice need?
CosyVoice does not publish an official RAM figure, so we do not state one. No CPU, RAM, GPU memory or disk figures are published. Install uses a Conda environment with Python 3.10; the optional ttsfrd text-normalisation wheel is Linux x86_64 / Python 3.10 only.
Does CosyVoice need a GPU?
Not stated in the README; the documented Docker server commands run with --runtime=nvidia.
How much does it cost to self-host CosyVoice?
The software is free to download. The project publishes no official RAM figure, so we do not estimate a monthly server cost.
Is CosyVoice free for commercial use?
CosyVoice is licensed under Apache-2.0 (code; the Fun-CosyVoice3 and CosyVoice2 model cards also list Apache-2.0) (Permissive). Internal business use is allowed under the license text. Free for internal and commercial use, modification and redistribution. Keep the license and notices. Check each model's card for its license before commercial use. This is a summary, not legal advice.
What are common problems when self-hosting CosyVoice?
There is no prebuilt image: you build the Docker image from runtime/python, and the documented server commands use the older CosyVoice-300M model, while the README recommends Fun-CosyVoice3-0.5B for best quality. Change --model_dir accordingly. Models are downloaded separately from ModelScope or Hugging Face into pretrained_models/ before use. vLLM acceleration works only with vLLM 0.9.0 or 0.11.x+ (0.10.x is untested, older is unsupported); the README suggests a separate Conda env because vLLM has strict dependency requirements. The repository moved from FunAudioLLM/CosyVoice to QwenAudio/CosyVoice; old links redirect.
Is CosyVoice still maintained?
Yes, it is actively developed. Last commit: May 25, 2026. Checked Nov 2, 2026.
How do I install CosyVoice?
The official method is Docker (self-built image, FastAPI server). The exact commands are in the "How to deploy" section above and in the official instructions: https://github.com/QwenAudio/CosyVoice#install
What are self-hosted alternatives to CosyVoice?
Other apps in Speech & translation: whisper.cpp, PDFMathTranslate, Fish Speech, Meetily, LibreTranslate, Anarlog, Speaches, Whisper ASR Webservice, Scriberr.
Sources
- GitHub repository, Gotcha 4: github.com/QwenAudio/CosyVoice
- Official documentation: github.com/QwenAudio/CosyVoice#readme
- License text: github.com/QwenAudio/CosyVoice/blob/main/LICENSE
- Hardware requirements, Deploy instructions, Gotcha 1, Gotcha 2, Gotcha 3: github.com/QwenAudio/CosyVoice#install
- Hetzner Cloud pricing (cost estimate): www.hetzner.com/cloud/regular-performance
- GitHub stars, commits and releases: GitHub REST API, fetched
Facts on this page were checked against the project's own repository and docs on 2026-11-02. Spotted something out of date? Tell us.