Ultimate Vocal Remover UVR

Ultimate Vocal Remover (UVR) is a free, open-source desktop app that uses AI models to split a song into its vocals and its instrumental, or into separate stems like drums and bass. It runs on Windows, Mac and Linux, it processes everything on your own computer, and the official download lives on the developers' GitHub page (the "Download" button on this page takes you straight there). No account, no subscription, no online version. Is it worth it? If you have a reasonably modern computer, on paper it's the most complete free option for making a karaoke track or an a cappella: free, local and full of models.

What UVR is, and what it isn't

UVR is a graphical front end for several families of source separation models, the kind of AI that learns to tell a voice apart from a guitar in a mixed track. It's built by two core developers, Anjok07 and aufr33, who say they trained all of the models in the package themselves, except the Demucs v3 and v4 4-stem models. The code is under the MIT license, which means free to use and modify, for anyone.

What it isn't: a website. The official project doesn't offer an online version or a mobile app: just installers for Windows and macOS and instructions for Linux. Anything with the same name on an app store or behind a login page isn't part of the official project.

How to remove vocals with UVR, step by step

Once it's installed, the main window is one screen. It looks busy, but for a first run you only need a handful of controls. These are the real names you'll see:

1. Click Select Input and pick your song. The Windows and Mac installers bundle everything it needs; if you run it from the source code, you'll need FFmpeg for anything that isn't a WAV file.
2. Click Select Output and choose the folder for the results.
3. Choose the output format: WAV, FLAC or MP3.
4. Under Choose Process Method, pick the model family. MDX-Net is the one shown in the official screenshot, and a good place to start.
5. Pick the specific model in the drop-down below it (in the official screenshot, MDX23C-InstVoc HQ).
6. Check GPU Conversion if you have a supported graphics card (the README sets the bar at an Nvidia GTX 1060 with 6 GB).
7. If you only want one part, check Vocals Only or Instrumental Only.
8. Press Start Processing and let it work.

Two more things worth knowing. Sample Mode (30s) processes just a 30-second slice, which is handy for comparing models before you commit a whole album. There's also a Select Saved Settings menu, and the README notes that the app remembers your settings when you close it. If you run into memory errors, the README's advice is to lower the Segment or Window size.

Which model to choose: MDX-Net, Demucs or VR Architecture

This is where UVR stops being a vocal remover and turns into a bit of a lab. The official documentation doesn't crown a "best model"; it credits where each family comes from, and that already tells you a lot:

• MDX-Net: built on the original code by Kuielab and Woosung Choi, and home to the newer MDX23C models trained by ZFTurbo. It's the method selected in the official screenshot, paired with the MDX23C-InstVoc HQ model, which is built for the vocals-versus-instrumental split. On a Mac, it's also one of the families with GPU acceleration.
• Demucs: the models from the Demucs project (facebookresearch/demucs). UVR supports versions 1 to 4, and the 4-stem models split a song into more than two parts, so it's the one to try when you want drums, bass and the rest as separate tracks.
• VR Architecture: based on the original VR Architecture code by tsurumeso. It's the third option in the menu, and one more to compare in Sample Mode.

Beyond those, you can download more models from inside the app, through the Settings menu (the wrench button next to Start Processing) and its Download Center. There's also a Roformer beta build that adds Roformer, SCnet and Bandit models; the official page offers it for Windows and for both kinds of Mac. My advice: pick one song you know well, run it through Sample Mode with two or three models, and trust your ears. Every song is different, which is exactly why Sample Mode exists.

Requirements: GPU, Windows and Linux

UVR runs its models on your computer, and these models are heavy. The README is honest about it:

• Nvidia GTX 1060 with 6 GB is the minimum for GPU processing, and 8 GB of video memory is recommended.
• AMD Radeon support is limited. If you have an AMD Radeon or Intel Arc card on Windows, the project points you to the beta build with DirectML.
• 64-bit only. Windows 10 or later is the target; Windows 7, and Intel Pentium or Celeron processors, aren't guaranteed to work.
• On Windows, install it on your main C: drive. The developers warn that a secondary drive causes instability.

Without a supported GPU, conversion runs on the processor, and the README warns that conversion times depend heavily on your hardware. The Windows installer is big (well over 1 GB), because it bundles Python, PyTorch and everything else, so you don't have to install anything first.

On Linux there's no installer: you download the code and run it with Python, following the official instructions for Debian and Arch-based systems. Fair warning: if "virtual environment" means nothing to you, stick to Windows or Mac.

Ultimate Vocal Remover on Mac

There are two Mac downloads, and picking the right one matters: one for Apple Silicon (listed as M1, arm64) and one for Intel (x86_64). The Mac version needs macOS Big Sur or later, and on Apple Silicon it can use the GPU (MPS) with Demucs v4 and all the MDX-Net models.

Two warnings from the official page, so they don't catch you off guard. The first launch can take 5 to 10 minutes, so don't assume it's frozen. And because of Apple's strict app security, macOS may refuse to open it; the README gives two Terminal commands to bypass that check (notarization), and recommends switching the security setting back on once UVR opens. If typing sudo makes you nervous, that's a perfectly sane reaction, and a sign an online tool may suit you better.

Is Ultimate Vocal Remover safe? Where to download it

The name is popular, so there are plenty of lookalikes. The safe rule is simple: download it only from the official GitHub repository by Anjok07, from its Releases page (the README also lists a Mediafire mirror for the same files). That's the page the button on this review opens. The source code is public and MIT-licensed, the installers are published on the project's own Releases page, and it doesn't ask for an account.

A note on dates: the latest stable release is v5.6, from 2023, and the Roformer beta builds were last updated in early 2025. The app works, but don't expect frequent updates.

When to use something else

UVR is the right call when you want quality, control and privacy for free, and you have the hardware. Look elsewhere if:

• You're on a phone or a Chromebook. UVR is desktop only.
• Your computer is old or has no dedicated graphics card and you have many songs to process. Conversion times depend heavily on your hardware.
• You want zero setup. A paid online stem splitter will get you a result in a browser tab, at the cost of uploading your music, and often a subscription.
• Your problem is noise, not music. If you want to clean up a voice recording or a podcast, a tool like AudioCleaner AI is built for that job.

For everything else, from a karaoke night to sampling a drum break, UVR is free, runs on your own machine, and gives you more models than you'll ever need. Not bad for an app that costs nothing. 🎤