Seed-VC Voice Conversion

Seed-VC: Zero-shot, Real-time Voice & Singing Voice Conversion Tool 🎙️🎶

Seed-VC is an advanced Voice Conversion (VC) and Singing Voice Conversion (SVC) project. Its standout features are zero-shot learning and real-time processing, allowing it to clone a target voice from just 1 to 30 seconds of reference audio without any training.

✨ Key Features

• Zero-shot Voice Conversion: Upload source and target reference audio for direct voice conversion.
• Zero-shot Real-time Voice Conversion: Supports ~300ms algorithm delay + ~100ms device side delay, suitable for online meetings, gaming, and live streaming.
• Zero-shot Singing Voice Conversion: Specifically optimized for singing, supporting higher sampling rates (44.1kHz) and high-pitched voice quality.
• Fine-tuning Support: Allows quick fine-tuning on custom data (minimum 1 utterance per speaker, as few as 100 steps).

🤖 Model Versions

Four pretrained models are provided for different scenarios:

• v1.0 tiny (seed-uvit-tat-xlsr-tiny): 25M params, 22050Hz, best for real-time VC.
• v1.0 small (seed-uvit-whisper-small-wavenet): 98M params, 22050Hz, best for offline high-quality VC.
• v1.0 base (seed-uvit-whisper-base): 200M params, 44100Hz, designed for singing voice conversion (SVC) with strong zero-shot performance.
• v2.0 (hubert-bsqvae-small): 67M+90M params, best at suppressing source speaker traits, supports voice & accent conversion.

🛠️ Usage Methods

Multiple interfaces are available:

• Command Line: Use inference.py or inference_v2.py for quick file conversion.
• Web UI: Run app_vc.py (VC), app_svc.py (SVC), or the integrated app.py for browser-based operation.
• Real-time GUI: Run real-time-gui.py for a graphical interface with adjustable parameters, suitable for virtual microphone routing.

💻 Installation & Requirements

Python 3.10 is recommended. Supports Windows, Linux, and Mac M-series chips. Install dependencies via pip to get started.

📈 Project Status

This project has 3.8k Stars and 492 Forks on GitHub, primarily written in Python (97.4%). It has an active community and is continuously updated.