Retrieval based Voice Conversion WebUI

Retrieval-based-Voice-Conversion-WebUI

A simple and easy-to-use voice changing framework based on VITS

Update log | FAQ | AutoDL·5 cents to train AI singer | Controlled experiment record | Online demonstration

English |中文简体|日本语| 한국어 ( Korean ) | Français | Türkçe | Português

The base model is trained using nearly 50 hours of open source high-quality VCTK training set. There are no copyright concerns. Please feel free to use it.

Please look forward to the bottom model of RVCv3, which has larger parameters, larger data, better results, basically the same inference speed, and requires less training data.

Training inference interface	Real-time voice changing interface
go-web.bat	go-realtime-gui.bat
You are free to choose what you want to do.	We have achieved end-to-end latency of 170ms. If you use ASIO input and output devices, you can achieve end-to-end 90ms latency, but it relies heavily on hardware driver support.

Introduction

This warehouse has the following characteristics

• Use top1 search to replace input source features with training set features to prevent timbre leakage

• Fast training even on relatively poor graphics cards

• Using a small amount of data for training can also yield better results (it is recommended to collect at least 10 minutes of low-noise speech data)

• The timbre can be changed through model fusion (with the help of ckpt-merge in the ckpt processing tab)

• Simple and easy to use web interface

• UVR5 model can be called to quickly separate vocals and accompaniment

• Use the most advanced human voice pitch extraction algorithm InterSpeech2023-RMVPE to eliminate the problem of mute sounds. Works best (significantly) but is faster and smaller than crepe_full

• A card I card acceleration support

Click here to view our demo video!

Environment configuration

The following instructions need to be executed in an environment with Python version greater than 3.8.

Common methods for Windows/Linux/MacOS and other platforms

Choose one of the following methods.

1. Install dependencies through pip

1. Install Pytorch and its core dependencies, skip if already installed. Reference from: https://pytorch.org/get-started/locally/

 pip install torch torchvision torchaudio

2. If it is a win system + Nvidia Ampere architecture (RTX30xx), according to the experience of #21, you need to specify the cuda version corresponding to pytorch

 pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117

3. Install the corresponding dependencies according to your own graphics card

• N card

 pip install -r requirements.txt

• A card/I card

 pip install -r requirements-dml.txt

• A card ROCM (Linux)

 pip install -r requirements-amd.txt

• I-cardIPEX(Linux)

 pip install -r requirements-ipex.txt

2. Install dependencies through poetry

Install Poetry dependency management tool, skip if it is already installed. Reference from: https://python-poetry.org/docs/#installation

 curl -sSL https://install.python-poetry.org | python3 -

When installing dependencies through Poetry, python is recommended to use versions 3.7-3.10. Other versions will conflict when installing llvmlite==0.39.0

 poetry init -n
poetry env use "path to your python.exe"poetry run pip install -r requirements.txt

MacOS

Dependencies can be installed through run.sh

 sh ./run.sh

Other pre-model preparation

RVC requires some other pre-models for inference and training.

You can download these models from our Hugging Face space.

1. Download assets

Below is a list with the names of all pre-models and other files required for RVC. You can find the scripts to download them in the tools folder.

• ./assets/hubert/hubert_base.pt

• ./assets/pretrained

• ./assets/uvr5_weights

If you want to use the v2 version model, you need to download it additionally.

• ./assets/pretrained_v2

2. Install ffmpeg

If ffmpeg and ffprobe are already installed, skip them.

Ubuntu/Debian users

 sudo apt install ffmpeg

MacOS users

 brew install ffmpeg

Windows users

After downloading, place it in the root directory.

• Download ffmpeg.exe

• Download ffprobe.exe

3. Download the files required for the rmvpe vocal pitch extraction algorithm

If you want to use the latest RMVPE vocal pitch extraction algorithm, you need to download the pitch extraction model parameters and place them in the RVC root directory.

• downloadrmvpe.pt

Download the dml environment of rmvpe (optional, A card/I card users)

• Downloadrmvpe.onnx

4. AMD graphics card Rocm (optional, Linux only)

If you want to run RVC on a Linux system based on AMD's Rocm technology, please install the required drivers here first.

If you are using Arch Linux, you can use pacman to install the required drivers:

pacman -S rocm-hip-sdk rocm-opencl-sdk

For some models of graphics cards, you may need to additionally configure the following environment variables (for example: RX6700XT):

export ROCM_PATH=/opt/rocm
export HSA_OVERRIDE_GFX_VERSION=10.3.0

Also make sure your current user is in render and video user groups:

sudo usermod -aG render $USERNAME
sudo usermod -aG video $USERNAME

Get started

direct start

Use the following command to start WebUI

 python infer-web.py

If you previously used Poetry to install dependencies, you can start WebUI in the following way

 poetry run python infer-web.py

Use integration packages

Download and unzip RVC-beta.7z

Windows users

Double-click go-web.bat

MacOS users

 sh ./run.sh

For I-card users who need to use IPEX technology (Linux only)

 source /opt/intel/oneapi/setvars.sh