rvc-boss/gpt-sovits

GPT-SoVITS is a voice cloning and text-to-speech system that can create a realistic copy of any voice from as little as one minute of audio, and in some cases produces usable results from just five seconds of a sample. The problem it solves is that traditional text-to-speech systems require recording hours of audio from a speaker to create a custom voice, making personalized voice synthesis accessible only to large production studios. GPT-SoVITS dramatically reduces this requirement to a practical minimum. The system works in two modes. In zero-shot mode, you provide a five-second reference audio clip and it immediately generates speech in that voice without any additional training. In few-shot mode, you provide about one minute of recordings and fine-tune the model to achieve better voice similarity and naturalness. The technology combines a GPT language model with the SoVITS voice synthesis framework, which is why the project has that name. It supports generating speech in multiple languages including English, Japanese, Korean, Cantonese, and Chinese, even when the voice training data was recorded in a different language. The project provides a web-based user interface built with Gradio, accessible through a browser, which includes built-in tools for separating vocals from background music, automatically segmenting recordings into training data, and labeling text transcripts. The tech stack is Python using PyTorch, and it runs on NVIDIA GPUs, AMD GPUs via ROCM, Apple Silicon, and standard CPUs. Windows users can download a pre-packaged version that requires minimal setup. You would use GPT-SoVITS for content creation, voiceover production, building interactive AI assistants with custom voices, or any application that needs high-quality personalized speech synthesis.