Learn — By WATMM Desk —
Is Ultimate Vocal Remover (UVR5) the best free stem separator?
UVR5 is the best free stem separator if you have a capable GPU and want to keep audio on your own machine, because it is a free, MIT-licensed front end for many open models, not one model. It is not automatically the most accurate: public leaderboards are now topped by newer band-split models, and the free web service MVSEP runs many of them for you.
Key facts
- Ultimate Vocal Remover GUI is MIT-licensed, free, and runs Demucs, MDX-Net and VR architecture models locally on Windows, macOS and Linux.
- Its README names an Nvidia GTX 1060 6GB as the minimum for GPU conversion and recommends 8 GB of VRAM.
- Demucs v4 (fine-tuned) scores 9.0 dB SDR on the MUSDB HQ test set, per its README; the repository is no longer maintained.
- On the MVSEP Multisong leaderboard, sorted by vocals, the top separation models are BS-RoFormer variants at about 12.3 dB vocal SDR.
- MVSEP is free for up to 50 separations a day, with files up to 10 minutes and 100 MB.
Short answer: for most people who want free, private, offline stem separation, yes, Ultimate Vocal Remover (UVR5) is the best place to start. But it is a free app that runs other people's models, not a model itself, and the most accurate models on public benchmarks today are newer than most of what people run in it by default.
What is free, and the limits
- UVR5 itself is "100% open-source and free," and its code is MIT-licensed, according to the project README. Installers cover Windows 10 and up and macOS Big Sur and up (Apple silicon and Intel builds), with manual steps for Debian- and Arch-based Linux. It only runs on 64-bit systems.
- The catch is hardware. The same README calls the models "computationally intensive," names an Nvidia GTX 1060 6GB as the minimum for GPU conversion, recommends 8 GB of VRAM, and says AMD support is limited. You can run on a CPU; expect it to be slow.
- Free without a GPU: MVSEP lets you do up to 50 separations a day free, with files up to 10 minutes and 100 MB. Without an account you get MP3 output only and a CAPTCHA; a free account adds WAV, FLAC and M4A output, with WAV and FLAC at 16-bit. Your audio leaves your machine, so check that before uploading unreleased work.
- Free and scriptable: Demucs is MIT-licensed and runs from the command line, but the repository says it is "not maintained anymore."
What UVR5 actually is
UVR5 is a desktop interface around three families of models: Demucs, MDX-Net and the VR architecture. Its README says the core developers trained the bundled models except the Demucs v3 and v4 four-stem ones, and that the MDX23C models were trained by ZFTurbo. So when someone says "UVR sounds better than X," the real question is which model inside UVR they used. Two users of the same app can get very different results.
How it compares on published benchmarks
There are two benchmarks worth knowing, and they are not interchangeable.
| Benchmark | What it measures | What it shows |
|---|---|---|
| MUSDB HQ test set | Average SDR across drums, bass, vocals, other | Demucs v4 fine-tuned scores 9.0 dB, against 5.9 for Spleeter in the same table (Demucs README) |
| MVSEP Multisong leaderboard | SDR on 100 songs across many genres, per stem | Top vocal entries are BS-RoFormer variants at about 12.3 dB (leaderboard, read 29 Oct 2026) |
SDR (signal-to-distortion ratio) is the standard score: higher means the separated stem is closer to the real one. MVSEP says it calculates it the same way as the Music Demixing Challenge (quality checker). The 2023 Sound Demixing Challenge paper reports that its best system improved SDR by more than 1.6 dB over the 2021 winner (arXiv), which gives a sense of how fast the field moved after Demucs v4.
Two cautions. First, you can't subtract a MUSDB score from a Multisong score; they are different songs and different averages. Second, the top four rows of the Multisong table are the original stems run through file formats (WAV, MP3, M4A), a reference ceiling, not models. Skip past them.
What we think
UVR5 wins on the things a free tool should win on: no account, no daily cap, your audio never leaves your machine, and you can swap in newer community models as they appear. It loses if you have an old laptop with no usable GPU, or if you want the current leaderboard leader without managing model files yourself; there, MVSEP's free tier is the easier route. The honest framing is not "UVR5 versus the rest" but "local versus hosted, and which model." Pick the model by the stem you need (a vocal-specialised model for acapellas, a four-stem model for drums and bass), and judge by ear on your own material, because a leaderboard average hides the song that matters to you.
Related
- Best AI stem splitters in 2026, our desk review of paid and free options
- How AI stem separation works, including why artefacts happen
- What are stems?
- Stem separation for DJ edits
Separating stems from a song you don't own doesn't give you the right to release what you make with them; clear the rights first.
Sources
- Ultimate Vocal Remover GUI README, GitHub (Anjok07) (primary source). Accessed 29 Oct 2026.
- facebookresearch/demucs README, GitHub (Meta Research) (primary source). Accessed 29 Oct 2026.
- Multisong leaderboard (sorted by vocals), MVSEP (primary source). Accessed 29 Oct 2026.
- Quality checker, MVSEP. Accessed 29 Oct 2026.
- MVSEP plans, MVSEP (primary source). Accessed 29 Oct 2026.
- The Sound Demixing Challenge 2023 – Music Demixing Track, arXiv (primary source). Accessed 29 Oct 2026.
Researched with AI assistance, written and checked by WATMM Desk.