GPT-SoVITS-RS
Turn GPT-SoVITS voices into a local HTTP service, without a Python runtime.
For people who already use GPT-SoVITS and want their assistant or application to
speak with a configured voice. Conversion and inference run in Rust. Training stays
in upstream GPT-SoVITS.
Try it | Listen | 中文上手 |
Use your models | API
Samples
Listen to Sun / download the generated WAV:
你好，欢迎回来。今天有什么想和我聊聊的吗？
Actual CPU inference with the maintainer's local Sun v2Pro + SV voice. Only generated
audio is shared, not the model, SV embedding, or reference. Settings and provenance.
This short example is not a quality or speed benchmark; long text can still lose words.
There is also a service confirmation sample and a
browser listening page.
The downloadable first-run demo below uses standard v2 weights, not Sun.
Quick Start
Linux x86_64, Docker Engine + Compose v2, Bash, curl, and SHA-256 tools.
Allow at least 6 GiB free disk and 4 GiB available memory. Downloads are about
1.1 GB plus the container; conversion and the first model load take time.
No Rust, Python, API key, or GPU is needed for this CPU demo.
The script downloads weights directly from their publisher, verifies checksums,
converts all four models inside the published 1.2.1 image, and checks the setup.
Read what it downloads and
the script before running it. Model licenses remain
separate from this project's MIT license.
git clone https://github.com/ricardomlee/gpt-sovits-rs.git
cd gpt-sovits-rs
bash examples/first-run/prepare.sh "$HOME/gpt-sovits-demo"
cd "$HOME/gpt-sovits-demo"
docker compose up -d --wait --wait-timeout 900
curl --fail-with-body --max-time 300 http://127.0.0.1:9881/tts \
-H 'Content-Type: application/json' --data-binary @request.json \
--output first.wav
Play first.wav with your usual audio player. The demo uses its own directory,
Compose project, and localhost-only port 9881, without changing a service on 9880.
Stop it with docker compose down in the demo directory; downloaded files stay there.
If preparation is interrupted, rerun the same command: verified downloads are reused.
Use a new directory, not an existing models or deployment directory.
Already have weights or use another platform?
Connect Your App
Once a voice is configured, send only its name and the text:
curl --fail-with-body http://127.0.0.1:9881/tts \
-H 'Content-Type: application/json' \
-d '{"voice":"demo","text":"The backup is complete."}' --output reply.wav
An OpenAI-compatible speech endpoint, sentence streaming, and multiple
voice profiles are also available. For an LLM assistant, submit completed short
sentences in order; agent integration covers playback
and errors. The service is not an authenticated public API.
Before You Choose It
- Supports compatible v2 and v2Pro weights, not arbitrary newer GPT-SoVITS checkpoints.
- A clean 3-10 second reference clip and matching transcript are still required.
- v2Pro needs an SV embedding for the full voice-conditioning path. The released converter
can convert an existing training embedding; it does not extract one from audio.
- Sentence splitting helps, but does not guarantee that every word will be spoken.
- No claim of being faster than Python. See measured performance.
- Models are not bundled or rehosted. Only use voices and recordings you have permission to use.
Help and Contribute
If the first request fails, check docker compose logs --tail 100 in the demo directory.
First-run troubleshooting covers downloads, memory, ports,
and readiness. Tell us where you got stuck, or what worked.
Do not attach private models, recordings, or credentials.
Deployment | API |
Development | Product goal
License and Credits
Code: MIT. Models keep their publishers' licenses.
The public sample has separate provenance.
Built on GPT-SoVITS and
Hugging Face Candle.
