Guides
Audio Isolating
UVR

Audio isolating with UVR5

Ultimate vocal remover (UVR) is an Free and Open Source AI audio editing tool with the ability to remove stems and other objects from a base audio. It is possible to execute the program through several ways, this section will explain how to use each of them.

Alternative 1: Locally execution

Download UVR5

To start using uvr locally, download and install the program via https://ultimatevocalremover.com/ (opens in a new tab)

ℹ️

When running the installer, Windows may mistakenly detect it as malware due to the lack of a license. If you have any concerns, you can review the program's source code or use alternative installation methods.

Download Models

When inside UVR, you first need to download separation models.

  • Click on the 🔧 icon.
  • Go to “Download Center”.
  • Choose your preferred models.

In the section you will find models structures, these are one of the best known:

RoFormer utilizes a Band-Split module to convert the input complex spectrogram into representations at the subband level (currently the best).

The RoFormer method contains two pretrained models, BS-RoFormer and Mel-Band RoFormer, which contain similar but not same structures, currently the best versions of these models are:

  • BS-Roformer-Viperx-1297: The best to separate vocals and instrumental.
  • Mel-Roformer-Viperx-1143: Separate vocals and instrumental.

For higher quality, use custom models

Custom Models

For information on how to use these models in the UVR, go to Custom Models Information

If you want to use these models, you need to download the following patch (opens in a new tab).

Make the Audio conversion

When the models have been downloaded, you can go back to the home screen to make your separation.

  • Select the desired structure and the downloaded model.
  • Select your audio, it can be easily downloaded from sites such as Cobalt.Tools (opens in a new tab)
  • Select the output path for the results and click convert.
Other Options
  • Segment Size: Larger consumes more resources, but may give better results.
  • Overlap: Amount of overlap between prediction windows.
  • Window Size: Change the audio processing bucket, a lower number indicates a much slower processing but a better worked result, and vice versa with a higher number.
  • Aggression Settings: Change the aggressiveness with which the separation will be performed, a higher number can eat important frequencies of the audio.

(This small application applies to any alternative that runs the GUI version of the program.)

Alternative 2: MVSEP

MVSEP is a website belonging to the UVR team which offers a fast and direct way to use the software in the cloud with a free version and paid upgrades.

To use mvsep, it is as easy as enter to their website (https://mvsep.com/ (opens in a new tab)) and registering. Apart from the mentioned models, MVSEP has some exclusive models, which are:

  • BSRoformer (ver 2024.08) or MelBand Roformer (ver 2024.08): The best models to separate instrumental and vocals.
  • Bandit v2: Remove SFX, speech, music, effects.
  • DeNoise by aufr23 (Standard & Aggressive): Remove noise and some artifacts.

Alternative 3: Google Collab

This is a list of currently functional uvr notebooks.

These notebooks are the main and most used ones, if you have any doubt about them or want to see if there are any editions to the notebooks it is recommended to check the official UVR server. (opens in a new tab)

Alternative 4: Kaggle

This is a list of currently functional uvr notebooks.

Custom Models Information

Important: You need to use overlap 2 on these models

Step 1: Place models in the correct folder

To use these models, you need to place the .ckpt and .yaml files in the following folders:

          • model.ckpt
              • config.yaml
  • Step 2: Select the model in UVR

    After placing the files in the folders, in UVR, select the model, this will open a window asking you to select the configuration file, select the model's .yaml file and check Roformer Model