README.md

---
license: mit
tags:
  - rvc
  - voice-conversion
  - real-time
  - audio
  - windows
  - cuda
---

![スクリーンショット](https://cdn-uploads.huggingface.co/production/uploads/6304d38abad6ce7fc027d0ce/APVdCbesIPoA0t83LvozM.png)
# RVC-Realtime-GUI
**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](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI).
It is designed for live microphone use, virtual audio routing, and audio-file
playback through a compact CustomTkinter interface.

Source code: [niel-blue/RVC-Realtime-GUI](https://github.com/niel-blue/RVC-Realtime-GUI)

## What's new in this release

**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

1. Extract the downloaded archive.
2. Open the `models` folder.
3. Start `RVC-Realtime-GUI-CUDA128.bat` for the CUDA 12.8 package, or
   `RVC-Realtime-GUI-CUDA118.bat` for the CUDA 11.8 package.
4. 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.

## Adding models

Create one folder per model under `models`:

```text
models/
  MyVoice/
    MyVoice.pth
    added_MyVoice.index
    preview.png
```

- 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.

## License and attribution

RVC-Realtime-GUI is based on
[RVC-Project/Retrieval-based-Voice-Conversion-WebUI](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI).
See the included `LICENSE` and `NOTICE.md` files, and the
[source repository](https://github.com/niel-blue/RVC-Realtime-GUI), for
attribution and license details.