Hysen LabsRetrieval-based-Voice-Conversion-WebUI
Easily train a good VC model with voice data <= 10 mins!
RVC WebUI: voice conversion trained on ten minutes of audio
A Python tool with a web interface for training voice conversion models, real time voice changing, and pitch handling that avoids the usual mute problems.
What the tool promises
The repository pitches itself as a tool to easily train a voice conversion model with voice data of ten minutes or less. The interface splits into a training and inference mode and a real time voice changing mode, and the user picks which operation to run. End to end latency is reported at 170 milliseconds, dropping to 90 with ASIO input and output devices, depending heavily on hardware driver support. The project is Python, MIT licensed.
Training tricks
Several techniques do the heavy lifting. Top1 retrieval replaces the input source features with training set features to prevent timbre leakage. Training stays fast even on relatively weak graphics cards, and small amounts of data still give usable results, with at least ten minutes of low noise voice data recommended. Model fusion can change timbre through the ckpt-merge option in the ckpt processing tab. The whole thing runs behind a simple web interface.
Pitch and separation
The tool can call pymss or MSST models to separate vocals and accompaniment quickly. For pitch, it uses the InterSpeech2023-RMVPE pitch extraction algorithm, described as state of the art, which eliminates mute issues while staying fast and using little resources. Those two capabilities cover the common preprocessing needs for voice data.
Hardware and dependencies
Dependency installation is organized by hardware. CPU, AMD, and Intel setups use a cpu requirements file, with DirectML available on Windows and CPU on Linux. NVIDIA RTX 50 series cards need CUDA 12.8 Torch first, while earlier NVIDIA cards use CUDA 11.8. The requirements files include download sources at the top: mainland China users can keep the default mirror, and official sources are a matter of replacing the index url flags while keeping package versions, CUDA suffixes, and the two stage order.
Editorial conclusion
RVC WebUI is a hands-on tool with explicit hardware guidance and a set of techniques aimed at getting clean voice conversion from small datasets. The README reads like a field guide rather than a spec sheet.
Community notes