gpt-sovits

# GPT-SoVITS A powerful few-shot voice conversion and text-to-speech WebUI by RVC-Boss. ## Overview GPT-SoVITS enables voice cloning with minimal training data. A 5-second vocal sample supports zero-shot TTS, while approximately 1 minute of data fine-tunes the model for improved voice similarity and realism. ## Key Features 1. **Zero-shot TTS** — Input a 5-second vocal sample for instant text-to-speech conversion. 2. **Few-shot TTS** — Fine-tune with ~1 minute of training data for higher fidelity. 3. **Cross-lingual Support** — Inference in languages different from the training dataset, supporting English, Japanese, Korean, Cantonese, and Chinese. 4. **WebUI Tools** — Integrated pipeline including: - Voice/accompaniment separation via UVR5 - Automatic training set segmentation - Multilingual ASR using Fun-ASR-Nano, SenseVoice, or classic FunASR - Text labeling for dataset creation ## Versions - **v2**: Added Korean and Cantonese support, expanded pretrained models from 2k to 5k hours, improved synthesis quality for low-quality reference audio. - **v3**: Higher timbre similarity with less training data, more stable GPT generation with fewer repetitions/omissions, richer emotional expression. - **v4**: Fixes metallic artifacts from non-integer upsampling in v3; natively outputs 48k audio instead of 24k. - **v2Pro**: Matches v4 performance with v2-level hardware requirements. ## Installation Supported platforms include Windows, Linux (CUDA/ROCm), macOS (MPS/CPU), and Docker. - Windows users can download an [integrated package](https://huggingface.co/lj1995/GPT-SoVITS-windows-package) and run `go-webui.bat`. - Conda-based install: `conda create -n GPTSoVits python=3.10 && conda activate GPTSoVits && bash install.sh` - Docker images are available on Docker Hub. - Colab notebooks are provided for cloud training. ## Pretrained Models Models are downloaded from Hugging Face and placed in `GPT_SoVITS/pretrained_models`. Additional models are needed for G2PW (Chinese text frontend), UVR5 (voice separation), and ASR backends (FunASR, Faster Whisper). ## Dataset Format Annotation files use the format: ``` vocal_path|speaker_name|language|text ``` Language codes: `zh` (Chinese), `ja` (Japanese), `en` (English), `ko` (Korean), `yue` (Cantonese). ## Usage - Launch WebUI: `python webui.py` or `python GPT_SoVITS/inference_webui.py` - Integrated package users: double-click `go-webui.bat` / `go-webui-v2.bat` - Command-line tools are available for UVR5 processing, audio segmentation, and ASR transcription. ## Inference Speed RTF benchmarks for GPT-SoVITS v2 ProPlus: 0.028 on RTX 4060 Ti, 0.014 on RTX 4090, 0.526 on Apple M4 CPU. An online demo is available on Hugging Face Spaces. ## Credits Built on research from VITS, SoundStorm, HiFi-GAN, BigVGAN, and other open-source speech synthesis projects. WebUI tools include UVR5, audio-slicer, FFmpeg, Gradio, and FunASR. ## License MIT License.