In this article
- How the settings were tested
- Results on a sparse mix
- Results on a dense rock mix
- Ultimate Vocal Remover best settings for an instrumental
- Best settings for a clean vocal
- Overlap, GPU Conversion and what they cost
- The five clicks
- Ultimate Vocal Remover models to download
- One result to distrust: VR Architecture on this build
- What this test does not cover
The Ultimate Vocal Remover best settings we measured are short: process method MDX-Net, model BS-Roformer-Viperx-1297, Segment Size Default, Overlap 2, GPU Conversion on. On two recordings whose real stems we have, that setup gave the cleanest vocal and the cleanest instrumental of seven models, and every slower setting we tried added time without adding quality.
| Goal | Process method | Model | Settings |
|---|---|---|---|
| Cleanest vocal or instrumental | MDX-Net | BS-Roformer-Viperx-1297 | Segment Default, Overlap 2, GPU on |
| Fast draft or a long batch | MDX-Net | MDX23C-InstVoc HQ | Segment Default, Overlap 2, GPU on |
| Drums, bass and other as separate files | Demucs | v4 | htdemucs | Defaults |
The rest of this page is the evidence: the scores, the run times, audio you can play, and the places where each model failed.
Part 1How the settings were tested
The program was Ultimate Vocal Remover v5.6.1, the January 2025 Roformer beta for Apple silicon, on a Mac with an Apple M5 Max chip and macOS 27. Output format was WAV. "Default" below means the value UVR shows when first opened.
Two excerpts were used, both from recordings published with their separate tracks, so each result can be compared with the real thing instead of judged by ear alone:
- A sparse mix. One minute of "Rainfall" by Liz Nelson and Jennifer Davies (MedleyDB, CC BY-SA 4.0): three voices, an acoustic guitar and an electric guitar.
- A dense mix. Forty seconds of "Karma" from the AlbumDB multitracks (CC BY 4.0): drums with cymbals, bass, guitars, piano and twelve vocal tracks, most of them backing vocals.
Each output is scored by SDR in decibels against the real track (the Vocal and Instr. columns below): higher is better, and every 6 dB halves the error. As a floor, handing over the untouched mix scores −1.9 dB as "the vocal" of the sparse excerpt. "Voice left" is how much of the real vocal can still be found in the instrumental file: −20 dB is a tenth of its level, −40 dB a hundredth.
Part 2Results on a sparse mix
| Model | Vocal | Instr. | Voice left | Time |
|---|---|---|---|---|
| BS-Roformer-Viperx-1297 | 19.7 | 21.7 | −45 | 38 s |
| MDX23C-InstVoc HQ | 18.4 | 20.3 | −38 | 10 s |
| Kim Vocal 2 | 16.8 | 18.7 | −35 | 4 s |
| UVR-MDX-NET Inst HQ 5 | 16.7 | 18.6 | −34 | 5 s |
| MB-Roformer-Inst-v1 | 16.4 | 18.4 | −32 | 21 s |
| Demucs v4 | htdemucs | 13.6 | 15.4 | −26 | 13 s |
| VR 1_HP-UVR | 11.5 | −4.5 | −20 | 6 s |
Times are UVR's own "Time Elapsed" for 60 seconds of audio. Demucs writes four stems; its instrumental here is its bass, drums and other files added together. BS-Roformer is first on every column except speed, and its lead is not small: 3 dB over the MDX-Net models is about 30% less error. MDX23C sits between them at a quarter of the time. The last row is not a typing mistake; it has its own section below.
The same 25 seconds, split by five models
Real output of Ultimate Vocal Remover v5.6.1 on an openly licensed recording, cut to this excerpt and not otherwise changed. Press play and switch versions: nothing to upload.
- Original mix10:25
- Vocals · BS-Roformer 129720:00 / 0:25
- Instrumental · BS-Roformer 129730:25
- Instrumental · MDX23C-InstVoc HQ40:25
- Instrumental · MDX-NET Inst HQ 550:25
- Vocals · Demucs v460:25
- Instrumental · VR 1_HP-UVR (hiss)70:25
- The real guitars, for reference80:25
Rainfall · Liz Nelson & Jennifer Davies · CC BY-SA 4.0. MedleyDB mix, 1:01 to 1:26. The separated files are output of Ultimate Vocal Remover v5.6.1, cut to this excerpt and not otherwise changed; shared under the same licence.

The run that scored best. UVR lists Roformer models under MDX-Net.
Part 3Results on a dense rock mix
| Model | Vocal | Instr. | Voice left | Time |
|---|---|---|---|---|
| BS-Roformer-Viperx-1297 | 7.5 | 14.9 | −15 | 25 s |
| MB-Roformer-Inst-v1 | 6.2 | 13.6 | −12 | 15 s |
| MDX23C-InstVoc HQ | 5.8 | 13.2 | −11 | 5 s |
| Kim Vocal 2 | 5.4 | 12.9 | −10 | 2 s |
| UVR-MDX-NET Inst HQ 5 | 5.2 | 12.6 | −10 | 3 s |
| Demucs v4 | htdemucs | 3.7 | 11.2 | −7 | 4 s |
Times here are for 40 seconds of audio. The order is almost the same and every score is far lower. The reason is in the "Voice left" column. Scored separately, the lead vocal is mostly gone from the BS-Roformer instrumental (−20 dB), but the backing vocals are only down to −11 dB, and with the MDX-Net models they are at −5 to −6 dB: about half their level, still singing in the "instrumental". Stacked harmonies behind a lead are the hardest thing for all of these models. If backing vocals are what you hear left over, a different overlap will not remove them.
Part 4Ultimate Vocal Remover best settings for an instrumental
Use the same model. It is tempting to pick one with "Inst" in its name, but on both excerpts BS-Roformer-Viperx-1297 made a better instrumental than UVR-MDX-NET Inst HQ 5 and MB-Roformer-Inst-v1: 3 dB better on the sparse mix, 1.3 to 2.3 dB on the dense one, with less voice left each time.
Cymbals were not the price. On the rock excerpt the energy above 8 kHz in every instrumental file was within 0.3 dB of the real instrumental's, so none of these six models dulled the top end to hide the voice. What they leave is voice, not holes.
Tick Instrumental Only if that is the only file you want. It saves the second write, not the processing time.
Part 5Best settings for a clean vocal
The vocal file fails the other way round. Very little music gets into it: on the sparse mix the guitars were 48 dB down in the BS-Roformer vocal and 39 dB down in the MDX-NET Inst HQ 5 one, and on the rock mix every model kept the band 33 to 35 dB down. What the vocal file lacks is part of the voice, the same backing vocals that stayed behind in the instrumental. So the model that leaves the least voice in the instrumental is also the one that gives the fullest vocal, and there is no reason to run two models. Tick Vocals Only to skip writing the other file.
Part 6Overlap, GPU Conversion and what they cost
| Setting changed | Score | Time for 60 s |
|---|---|---|
| BS-Roformer, Overlap 2 → 8 | −0.1 dB | 0:38 → 2:32 |
| Inst HQ 5, Overlap Default → 0.99 | +0.2 dB | 0:05 → 2:29 |
| BS-Roformer, GPU Conversion off | none | 0:38 → 2:49 |
| VR 1_HP-UVR, Aggression 5 → 10 | +0.1 dB | 0:06 → 0:04 |
Overlap is the setting most guides tell you to raise. Here it multiplied the run time by four with a Roformer model and by thirty with an MDX-Net model, for a change of a fifth of a decibel or less, in one case downward. Leave it at the default.
GPU Conversion changed nothing in the audio: the files scored the same to two decimals. It only changed the wait, by more than four times. Keep it on if your machine has a supported GPU.

Overlap 8: four times the wait for the same result.
Part 7The five clicks
- Select Input and choose the song. Select Output sets the folder the files go to.
- Under Choose Process Method pick MDX-Net.
- Under Choose MDX-Net Model pick BS-Roformer-Viperx-1297. Leave Segment Size on Default and Overlap on 2.
- Keep GPU Conversion ticked. Tick Vocals Only or Instrumental Only if you want one file.
- Press Start Processing. The log under the button ends with "Process complete" and the time it took.
UVR names the files after the song and the model: ours were 1_rainfall-60s_model_bs_roformer_ep_317_sdr_12.9755_(Vocals).wav and the same with (Instrumental). The long middle part is the model's file name, which is how you tell two runs apart later.
Part 8Ultimate Vocal Remover models to download
The app ships with three of the models above: UVR-MDX-NET Inst HQ 5, VR 1_HP-UVR and Demucs v4 htdemucs. The others come from the Download Center: press the wrench button, open the Download Center tab, pick the model under MDX-Net and press the download button. The files come from the project's model release; for this test we took them from there directly and put them in the app's MDX_Net_Models folder, which UVR reads the same way. BS-Roformer-Viperx-1297 is 639 MB, MB-Roformer-Inst-v1 913 MB, MDX23C-InstVoc HQ 448 MB and Kim Vocal 2 67 MB. If you download only one, make it the first.

The Download Center, behind the wrench button.
The program itself is on the official releases page.
Part 9One result to distrust: VR Architecture on this build
The VR Architecture model that ships with the app, 1_HP-UVR, wrote an instrumental louder than the mix it came from. Below 8 kHz the file matches the real guitars closely. Above 8 kHz it carries noise at −22 dB where the real instrumental has −71 dB. It is the last row of the player above; turn the volume down first.
The same noise, at the same level, was in the instrumental file of all four VR runs: both excerpts, GPU Conversion on and off, Aggression 5 and 10. In two of them (the rock excerpt, and the sparse one with the GPU off) the vocal file carried it as well. We tested one VR model on one Mac, so this may be particular to that model or to this build, but it is why VR is not in the table at the top.
Part 10What this test does not cover
Two excerpts and one computer are a small test. Windows with an NVIDIA card will give different times. Segment Size, Ensemble Mode, the denoise option and the other downloadable models were not run. Scores from two songs will not rank models on every song; they do show which settings are worth your time and which are not.
If you would rather not install anything, our vocal remover runs a model of the same BS-Roformer family in the cloud and returns the vocal and the instrumental; the stem splitter adds drums, bass and the rest. Both let you play the result before you download it, and neither needs a model file. For speech over background music in an interview or podcast, use the voice remover to isolate the spoken track.
Last updated Nov 16, 2026. The figures here come from our own measurements, written up on each tool's page. When a tool changes, this page is checked again.


