Clone Any Voice in 1 Minute with GPT-SoVITS

Conda keeps GPT-SoVITS's many dependencies isolated so they don't break anything else on your machine. This command creates a clean Python 3.10 environment and then activates it. Every command after this must be run inside that same activated environment.

This clones the GPT-SoVITS repository to your computer — it downloads all the scripts and config files you need. If you don't have Git, you can also download the ZIP from the GitHub page and unzip it, then navigate into the folder.

This single script installs all Python packages AND downloads the required pretrained AI models automatically. Pick the command that matches your hardware. CU126 means CUDA 12.6 (most common for recent NVIDIA cards). Use CPU if you have no NVIDIA GPU. The --source HF flag downloads from Hugging Face — if you are in China, swap HF for ModelScope.

This starts the local web server. After a few seconds it will print a URL like http://127.0.0.1:9874. Open that URL in any browser. The full interface loads there — you never need to touch code.

In the browser UI, go to the 'TTS Inference' tab. Upload your reference audio clip (WAV or MP3, 5–30 seconds works for zero-shot; 1 minute is better). Type the text you want spoken into the text box. Select the language of both the reference audio and the output text. Click 'Synthesize'. The result plays directly in the browser and can be downloaded. No training is needed for a quick zero-shot test — the model clones the voice on the spot.

Zero-shot is fast but fine-tuning gives noticeably better results. In the WebUI, go to the 'Training' tab. Upload your 1-minute audio, use the built-in tools to auto-slice it into segments and label them, then click the training buttons in order (SoVITS train, then GPT train). Training takes 20–60 minutes on a GPU. Once done, load your new model files in the Inference tab for higher-quality output. The WebUI walks you through each sub-step visually.