A self-contained Windows desktop application for low-latency, real-time RVC voice conversion.
This release is a focused real-time inference client derived from
RVC-WebUI.
It is designed for live microphone use, virtual audio routing, and audio-file
playback through a compact CustomTkinter interface.
This build includes the real-time inference performance updates introduced
after 2026-07-18.
CUDA Graph warm-up for lower per-chunk inference overhead after startup
Updated real-time inference path and input/output noise-reduction fixes
CUDA 12.8 runtime for current NVIDIA GPUs, including Blackwell-compatible
environments
Native ASIO buffer preservation: starts ASIO with the driver's preferred
buffer size instead of allowing the high-latency default to select a larger
buffer
Download
Open the Files and versions tab and download the archive that matches your
NVIDIA GPU.
Package
Status
Use it when
RVC-Realtime-GUI_cu128_260729.7z
Current standard build
You use a current NVIDIA GPU, including Blackwell-compatible environments.
RVC-Realtime-GUI_cu118_260729.7z
Legacy build (available)
You use an older NVIDIA GPU that cannot use the CUDA 12.8 build.
Extract the archive to a short local path, for example C:\RVC-Realtime-GUI.
Do not run it from inside the archive.
Quick start
Extract the downloaded archive.
Open the models folder.
Start RVC-Realtime-GUI-CUDA128.bat for the CUDA 12.8 package, or
RVC-Realtime-GUI-CUDA118.bat for the CUDA 11.8 package.
Choose a model, audio devices, and press Start.
The package already contains its Python runtime, PyTorch/CUDA stack, FFmpeg,
and inference assets. No separate Python, CUDA Toolkit, FFmpeg, or RVC-WebUI
installation is required.
The folder name is shown as the model name in the app.
Put the model .pth and its added_*.index file in the same folder.
An optional .png, .jpg, or .jpeg image in that folder is shown as the
model preview.
Restart the app or use Reload after adding or changing a model.
Features
Real-time RVC inference with CUDA Graph warm-up
WASAPI and native ASIO audio-device routing
Separate input, output, and monitor devices
Model gallery with previews and model-specific general settings
Audio-file input through bundled FFmpeg
WAV recording: separate input/output files, mixed recording, or split L/R
Japanese and English interface
Runtime log viewer, log export, and recording status display
Notes
This release is for Windows and NVIDIA GPUs.
Start with conservative chunk and buffer settings, then lower latency only
after confirming stable audio on your device chain.
Native ASIO uses the driver's preferred hardware buffer size; RVC processing
chunk length is kept separate from that setting.
ASIO and WASAPI can differ substantially by driver and virtual-audio setup.
RVC models are not interchangeable with this package unless they are valid
RVC .pth models. Use only voices and audio material you are authorized to
use.
