What is this?
RVC WebUI is a web-based tool for training and running voice conversion (voice-changing) models, letting you clone a target voice's timbre using as little as about 10 minutes of audio.
RVC WebUI is a web-based tool for training and running voice conversion (voice-changing) models, letting you clone a target voice's timbre using as little as about 10 minutes of audio.Is this a desktop app or something I open in a browser?It is a self-hosted Gradio web app: you launch it locally and it serves a WebUI (training/inference tab and a separate real-time voice-changing tab) that you open in a browser, listening on port 7865 by default.
It is a self-hosted Gradio web app: you launch it locally and it serves a WebUI (training/inference tab and a separate real-time voice-changing tab) that you open in a browser, listening on port 7865 by default.How do I get it running?Install Python 3.12 x64 and create a virtual environment, install the requirements file matching your hardware, download the pretrained models from Hugging Face into the project folders, then launch with go-webui.bat on Windows or the equivalent command on Linux; models load from assets/weights and assets/indices.
Install Python 3.12 x64 and create a virtual environment, install the requirements file matching your hardware, download the pretrained models from Hugging Face into the project folders, then launch with go-webui.bat on Windows or the equivalent command on Linux; models load from assets/weights and assets/indices.Do I need an NVIDIA GPU to use it?Not strictly: NVIDIA GPUs use CUDA 12.8 (RTX 50-series) or 11.8 (older RTX) Torch builds, but AMD/Intel hardware is supported via DirectML on Windows or plain CPU on Linux.
Not strictly: NVIDIA GPUs use CUDA 12.8 (RTX 50-series) or 11.8 (older RTX) Torch builds, but AMD/Intel hardware is supported via DirectML on Windows or plain CPU on Linux.How fast is the real-time voice changing?End-to-end latency is about 170ms by default, dropping to roughly 90ms with ASIO input/output audio devices, though the lower figure depends heavily on your hardware's driver support.
End-to-end latency is about 170ms by default, dropping to roughly 90ms with ASIO input/output audio devices, though the lower figure depends heavily on your hardware's driver support.How does it avoid the source speaker's voice leaking through into the output?It uses top-1 feature retrieval that swaps the input speaker's features for matched features from the training set, which prevents timbre leakage.
It uses top-1 feature retrieval that swaps the input speaker's features for matched features from the training set, which prevents timbre leakage.What pitch extraction method does it use?It uses RMVPE, an InterSpeech2023 algorithm chosen for being fast, low-resource, and resistant to the 'mute note' artifacts that plague earlier pitch trackers.
It uses RMVPE, an InterSpeech2023 algorithm chosen for being fast, low-resource, and resistant to the 'mute note' artifacts that plague earlier pitch trackers.Can I combine two trained voice models?Yes, the checkpoint-processing tab includes a ckpt-merge option that fuses two models to blend their timbres.
Yes, the checkpoint-processing tab includes a ckpt-merge option that fuses two models to blend their timbres.Can it pull just the vocals out of a song before conversion?Yes, it can call pymss/MSST models to quickly separate vocals from accompaniment music.
Yes, it can call pymss/MSST models to quickly separate vocals from accompaniment music.What are similar projects?so-vits-svc, GPT-SoVITS, voice-pro, fish-speech, personaplex, ChatTTS, moshi, DramaBox
so-vits-svc, GPT-SoVITS, voice-pro, fish-speech, personaplex, ChatTTS, moshi, DramaBoxWhich topics is it in?conversion, vits, voice-conversion, conversational-ai, model-training, converter, amd, voice-cloning, voice, gradio, audio-processing, webui
conversion, vits, voice-conversion, conversational-ai, model-training, converter, amd, voice-cloning, voice, gradio, audio-processing, webui