GPT-SoVITS is a voice cloning and text-to-speech application that requires only one minute of audio data for few-shot model fine-tuning and five seconds of audio for zero-shot inference. It supports cross-lingual generation in five languages and can be deployed locally via Docker, conda, or source on Windows, Linux, and macOS.
Project overview
It lowers the data barrier for training text-to-speech models, requiring only 1 minute of training data to produce a fine-tuned model, and supports cross-lingual inference in English, Japanese, Korean, Cantonese, and Chinese.
Project type
Audio & Speech
Use cases
Meeting Notes · Audio & Speech
Deployment
Python environment
License
MIT
Best for
Creators, developers, and general users seeking a self-hostable application for voice cloning and text-to-speech generation using minimal training data.
Key capabilities
Accepts a 5-second vocal sample to perform text-to-speech conversion without fine-tuning.
Allows fine-tuning the model with 1 minute of training data to improve voice similarity and realism.
Executes inference in languages different from the training dataset, supporting English, Japanese, Korean, Cantonese, and Chinese.
Development of pretrained finetuned preset GPT models for better emotion control is planned.
Capability to mix different TTS models is planned for future implementation.
Limitations and risks
Models trained with GPUs on Macs produce significantly lower quality results, requiring the use of CPUs instead on those devices.
Version 3 experienced an issue with metallic artifacts caused by non-integer multiple upsampling.
Getting started
Setup is rated as medium difficulty. It requires creating a conda environment (Python 3.10), activating it, running an install script (such as install.ps1 on Windows with specified device and source flags), and manually downloading pretrained models to place in the GPT_SoVITS/pretrained_models directory.
Alternatives and comparisons
A tokenizer-free text-to-speech model supporting 30 languages, 48kHz audio output, voice design from text descriptions, and accurate voice cloning.
A local-first voice I/O stack bridging voice cloning, multi-engine speech generation, and global dictation with agent voice output.
Converts e-books into audiobooks with chapters and metadata using text-to-speech technology, supporting multiple engines and 1158 languages.
GitHub project description: 1 min voice data can also be used to train a good TTS model! (few shot voice cloning)
README: 1. **Zero-shot TTS:** Input a 5-second vocal sample and experience instant text-to-speech conversion. 2. **Few-shot TTS:** Fine-tune the model with just 1 minute of training data…
README: [](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main…
Release: 20250606v2pro
README: Version 4 fixes the issue of metallic artifacts in Version 3 caused by non-integer multiple upsampling, and natively outputs 48k audio to prevent muffled sound (whereas Version 3…