GPT-SoVITS: Clone Voices and Generate Speech from Text

Overview

GPT-SoVITS is a self-hostable toolkit for creating text-to-speech voices and converting speech to a target voice. It is designed to make voice adaptation possible with little speaker data: users can try zero-shot synthesis from a short reference sample or fine-tune a model using a small dataset.

It is aimed at people who want an interactive workflow for preparing audio, training or fine-tuning voice models, and generating speech, rather than only a command-line inference library. Its WebUI brings together dataset preparation tools and multilingual speech workflows.

Key Features

• Zero-shot text-to-speech using a short voice reference sample.
• Few-shot fine-tuning for improved voice similarity using a small amount of training audio.
• Cross-lingual inference, with support listed for Chinese, English, Japanese, Korean, and Cantonese.
• WebUI workflows for audio segmentation, transcription, text labeling, and model fine-tuning.
• Optional vocal and accompaniment separation and reverberation removal through UVR5 tools.
• Multiple model generations and pretrained model options, with release notes describing their differences.
• Installation options for Windows, Linux, macOS, and Docker, with CPU and accelerator configurations documented.

Use Cases

• Creators can generate narration in a consistent voice without recording every line, when they have suitable reference audio and model setup.
• Voice actors or audio producers can prototype alternate-language speech using a speaker reference and supported language workflows.
• Researchers and hobbyists can experiment with few-shot voice adaptation and compare the project’s model versions.
• Users building a speech dataset can use the WebUI’s segmentation and ASR tools to prepare clips and transcripts before fine-tuning.

Project Facts

• Language: Python
• License: MIT
• Stars: 62.3k
• Forks: 6.7k
• Topics: text-to-speech, tts, vits, voice-clone, voice-cloneai, voice-cloning
• Archived: false

Getting Started

The README documents installation scripts for each supported platform. For example, on Linux:

conda create -n GPTSoVits python=3.10
conda activate GPTSoVits
bash install.sh --device <CU126|CU128|ROCM|CPU> --source <HF|HF-Mirror|ModelScope> [--download-uvr5]

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

Alternatives

• Spark-TTS: Spark-TTS uses a Qwen2.5-based model and focuses on Chinese-English synthesis and zero-shot voice cloning.
• VoxCPM: VoxCPM is a tokenizer-free multilingual system for speech generation, voice design, and cloning, with a 2B-parameter model option.
• Chatterbox: Chatterbox offers multiple TTS model variants, including multilingual synthesis and paralinguistic tags for expressive speech.
• CSM: CSM generates speech from text and audio context using a Llama backbone and an audio-code decoder.

Considerations

• Practical setup involves Python, PyTorch, FFmpeg, pretrained models, and potentially substantial storage and compute resources. The README lists CPU and accelerator paths, but does not make all workflows equally convenient on every device.
• macOS GPU training is noted as producing lower quality than training on other devices, so the README recommends CPU use for that platform.
• Voice quality and behavior depend on the reference audio, training data, selected model version, and hardware. The project offers several generations with differing characteristics.
• The repository has a large open issue count, so users should review current issues and setup guidance when troubleshooting.