GPT-SoVITS

GPT-SoVITS is a local voice cloning and text-to-speech tool for people creating speech in a custom voice. Its web interface covers speech generation, voice training and preparation of audio datasets.

• Speech from a reference: Generate speech using a five-second vocal sample without first training a custom voice model.
• Voice fine-tuning: Train with as little as a minute of recordings to improve similarity to the speaker.
• Multilingual speech: Produce English, Japanese, Korean, Cantonese and Chinese audio, including speech in a language different from the training recordings.
• Recording preparation: Separate vocals from accompaniment, remove reverberation, split recordings into training clips and proofread transcriptions and text labels.

The open-source code uses the MIT license. It runs on Windows and macOS, with Docker deployment also available. CPU inference and NVIDIA CUDA hardware are supported. On Macs, the project uses CPU training because GPU training can produce substantially lower voice quality.

Local use requires downloaded models. Transcription tools include Chinese ASR and Faster Whisper for English and Japanese. The Hugging Face demo and AutoDL Cloud option run on external infrastructure.