GPT-SoVITS: Clone Voices and Generate Speech from Text

Overview

GPT-SoVITS is a self-hosted toolkit for voice cloning and text-to-speech. It addresses the need to generate speech in a target voice without assembling a large training corpus: the project supports zero-shot synthesis from a short reference sample and fine-tuning from a small dataset.

Its WebUI brings together inference, model fine-tuning, and audio-dataset preparation. It is aimed at users who want to create or run a custom voice model, rather than teams seeking a ready-made hosted speech API.

Key Features

• Generate speech from text using a short reference voice sample, without fine-tuning.
• Fine-tune a voice model using a small speaker dataset.
• Support inference across Chinese, English, Japanese, Korean, and Cantonese.
• Provide WebUI workflows for audio segmentation, denoising, transcription, and text labeling.
• Offer command-line entry points for inference and dataset processing.
• Support multiple model generations, with different trade-offs in audio quality and model behavior.
• Include Docker Compose options alongside installation scripts for Windows, Linux, and macOS.

Use Cases

• Independent creators can produce narration in a consistent voice when they have a suitable voice sample and permission to use it.
• Researchers and developers can experiment with few-shot speech synthesis and compare model versions on their own data.
• Localization teams can prototype speech in supported languages using a reference voice, then assess whether the output meets their quality needs.
• Users building a custom voice dataset can use the included segmentation and transcription tools to prepare audio for fine-tuning.

Getting Started

The README recommends Python 3.10 and provides platform-specific installation scripts. For Linux, from the repository root:

conda create -n GPTSoVits python=3.10
conda activate GPTSoVits
bash install.sh --device <CU126|CU128|ROCM|CPU> --source <HF|HF-Mirror|ModelScope>

Then start the WebUI with python webui.py. See the README for platform-specific instructions, model downloads, and inference details.

Alternatives

• ChatTTS: ChatTTS focuses on Chinese and English dialogue speech with prosody controls, while GPT-SoVITS emphasizes voice cloning and speaker fine-tuning.
• index-tts-lora: index-tts-lora adds LoRA fine-tuning to IndexTTS, while GPT-SoVITS provides its own voice-cloning and TTS training and inference toolkit.
• open-notebooklm: open-notebooklm creates podcast dialogue from PDFs using an API, while GPT-SoVITS synthesizes speech from text using voice samples or fine-tuned speakers.
• chatterbox: Chatterbox offers a model family with multilingual and expressive speech options, while GPT-SoVITS centers on zero-shot voice cloning and small-dataset fine-tuning.

Considerations

• GPU acceleration is supported, but the README also documents CPU options. The project notes lower quality for models trained on Mac GPUs and recommends CPU training there.
• Installation involves pretrained model downloads and additional audio tools such as FFmpeg. ASR and vocal-separation models may require separate downloads.
• The README describes multiple model generations and upgrade steps, so users should check model compatibility and instructions for their chosen version.
• The project is active, with recent commits and contributions from many people, but it has a substantial open-issue backlog and relatively few issues closed recently. There are no reported tests in the repository health data.
• The MIT license permits broad use, subject to its terms. Users should also ensure they have the rights and consent needed for any voice recordings they use or generate.