GPT-SoVITS
RVC-Boss/GPT-SoVITSReviewed by AI
Checked by RepoPicks Radar on
What is GPT-SoVITS?
A voice cloning and text-to-speech workshop with a page in your browser. You give it a 5-second sample for an instant copy of a voice, or about one minute of audio to train a better one, and it then reads your text in that voice, including in a language the sample was not in. The page also helps you prepare the training audio: separating voice from music, cutting clips and writing out what is said. Windows users can download a ready package and double-click `go-webui.bat`; otherwise it is conda, Python and an Nvidia card, or Docker. There is a free online demo.
GPT-SoVITS is free and open source (MIT) and can replace ElevenLabs, Respeecher, and Murf.
How easy is it to install GPT-SoVITS?
A · download and open · Web
- Windows B
- macOS D
- Linux D
- Web A
The web is an A: the README links a free demo page where you type and listen, plus Colab notebooks for training. Windows is a B: you download the "integrated package" (a 7z file from Hugging Face), unpack it and double-click `go-webui.bat`. Linux and Mac are a D: you create a conda environment and run `bash install.sh --device <CU126|CU128|ROCM|MPS|CPU> --source <HF|HF-Mirror|ModelScope>`, which installs ffmpeg, cmake and a compiler and downloads several model packs; the full feature set assumes an Nvidia card, and Docker images need an Nvidia runtime.
GPT-SoVITS on GitHub
- Stars
- 64,720on Nov 10, 2026
- Downloads
- —all releases
- Growth
- —no gain in the last reading
Stars over time
Facts
- Runs on
- Windows, macOS, Linux, Web
- You need to install
- Python, Docker, NVIDIA graphics card, FFmpeg
- To build from source
- Python, Docker, NVIDIA graphics card, FFmpeg
- Disk space
- 450 MB–1.2 GB · estimate
- Accounts or keys
- None
- Cost
- Free · No price: MIT code, trained models published openly on Hugging Face and ModelScope, and a free online demo on Hugging Face.
- License
- MIT (Permissive)
- Can replace
- ElevenLabs, Respeecher, Murf
- Last change on GitHub
- Nov 9, 2026
GPT-SoVITS risk signals (6)
- Attentionvoice cloning from a short sample
README Features: "Zero-shot TTS: Input a 5-second vocal sample and experience instant text-to-speech conversion" and "Few-shot TTS: Fine-tune the model with just 1 minute of training data". Nothing in the README marks the output as machine-made, and copying a real person's voice without permission is restricted by law in several countries.
AI review
- Infomodels and images come from personal accounts
install.sh and install.ps1 download the model packs from "https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/..." (or hf-mirror.com / modelscope.cn), the Windows package comes from "huggingface.co/lj1995/GPT-SoVITS-windows-package", and docker-compose.yaml pulls the images "xxxxrt666/gpt-sovits:latest-cu126" and similar from Docker Hub — all personal accounts rather than the GitHub releases.
AI review
- Infodocker setup shares the folder and opens ports
docker-compose.yaml mounts the whole project folder (`- .:/workspace/GPT-SoVITS`) and publishes five ports (9871-9874 and 9880, the API). The README warns: "Docker Compose will mount all files in the current directory", and asks for `shm_size: 16g` of shared memory.
AI review
- Infomany versions with manual model downloads
The README has separate release notes and separate model downloads for v2, v3, v4, v2Pro and v5, each telling you to fetch files from Hugging Face into `GPT_SoVITS/pretrained_models`; V5 even says to "Clone the latest code" from a different branch (`cuda_graph_accel_v5`). The newest tagged release is 20250606v2pro, while the code keeps changing daily.
AI review
- Infomac results are worse
README macOS section: "The models trained with GPUs on Macs result in significantly lower quality compared to those trained on other devices, so we are temporarily using CPUs instead." The measured speed table lists 0.526 RTF on an M4 processor against 0.014 on a 4090 graphics card.
AI review
- Infomain guides are outside the repository
The user guide links point to yuque.com (Chinese) and rentry.co (English) rather than to documentation in the repository, and Chinese users are sent to a yuque page to download the models and the Windows package.
AI review
Evidence is shown in its original language.
What Radar could not check about GPT-SoVITS
- How good the cloned voice really sounds from 5 seconds or from one minute of audio
- Real disk space after the pretrained models, the speech-recognition models and the voice-separation models download
- Whether the Windows "integrated package" (dated 20250228) matches the current code and the newer versions
- Whether it runs usefully without an Nvidia card, and how slow it is
- Whether the free online demo stays available
- What is inside the third-party model packs and Docker images that the installer downloads