UVR Vocal Bleed: Why Ultimate Vocal Remover Leaves Instruments in the Vocals

By IntrectNovember 15, 2026Last reviewed StemsBleedUVRGuide

Ultimate Vocal Remover (UVR) is the free desktop app many people use to pull a vocal out of a song. It often does well, and then you solo the result and hear a snare under the verse, a hi-hat in the gaps, or a synth pad humming along with the long notes. That is bleed. This guide covers what UVR can change about it and what to do with the vocal afterwards.

We make de-leak, a plug-in for removing bleed from separated stems. UVR is not ours, and it is free; we say which steps need no extra software.

The short answer

Pick the model on a test passage, not by reputation, then clean what is left in your DAW. Compare two or three UVR models (or an ensemble) on the same 20 seconds, keep the vocal with the least bleed in the loud sections, gate the gaps between phrases, and use a per-band bleed remover for what plays under the voice. Listen to what each step removes.

Why UVR leaves instruments in the vocal

UVR is a graphical front end for several separation architectures. Its README credits the original VR Architecture code to tsurumeso, MDX-Net to Kuielab and Woosung Choi, and Demucs to its author (Adefossez on GitHub). It also says UVR’s core developers trained all of the models in the package except the Demucs v3 and v4 four-stem models. The version on the project page when we checked was 5.6.

Whatever the model, the job is the same: estimate the vocal from a finished stereo mix. Where a drum hit or a chord sits in the same frequencies as the singer at the same moment, the model has to split that energy, and some of it stays in the vocal. Source-separation research calls that interference; engineers call it bleed. It is different from the glassy or watery residue a separator can add, which our guide to shimmer covers.

Choices inside UVR that change the bleed

None of these are guarantees. They are the settings that change what ends up in the vocal, so they are worth testing before you clean anything.

A fair model test in ten minutes

  1. Cut a 20-second passage that has a quiet gap, a loud chorus line and a held note.
  2. Run it through two or three models (and one ensemble if you use them) with the same output format.
  3. Load the vocals side by side in your DAW and match their loudness.
  4. Listen for three things: drums or chords in the gap, instruments under the loudest line, and how full the voice still sounds. The cleanest gap is not always the best vocal; a model can buy silence by thinning the voice.
  5. Keep the best one and run the whole song through it.

Cleaning the vocal after UVR

  1. The gaps. Between phrases, bleed is the only thing on the track. A noise gate or simple region edits remove it there, free, in any DAW.
  2. One obvious band. A high-pass below the singer’s lowest note takes out kick and bass thump. Stop at one or two EQ moves; more starts removing the voice.
  3. Bleed under the voice. A level gate cannot help here, because it opens for the voice and lets the snare through with it. You need a tool that decides per frequency band whether the sound belongs to the vocal. de-leak does this with a learned 16-band Gate and a 64-band model called Leak Clean; it runs as VST3 or CLAP on macOS and Windows, and as an Audio Unit and AAX on macOS (AAX for native Pro Tools on Apple Silicon). Start with the Medium preset and the Gate engine, compare Leak Clean or Gate → Clean on the same passage, and keep what sounds right.
  4. Check what was removed. In de-leak, Residual plays the dry signal minus the processed signal. Turn your monitoring up by 20–40 dB and listen: drums, instruments and room mean it is working; words or melody mean it is too strong. Our guide on hearing what a plug-in removed shows how to do the same check with any tool.

de-leak reports a fixed delay of about a quarter of a second that your DAW compensates automatically, so it is for mixing and bouncing, not live monitoring. It has a 14-day trial; details are on the product page.

What cannot be fixed afterwards

Found bleed in a stem?

de-leak has a 14-day trial for macOS and Windows hosts that load VST3 or CLAP. Comparing tools first? Read the bleed comparison.

Related guides

Sources

FAQ

Why are there drums in my UVR vocal?

UVR estimates the vocal from a finished mix. Where drums or instruments share frequencies with the voice at the same moment, part of them stays in the vocal estimate. That is bleed. Different models leave different amounts and different kinds, but no model removes all of it on every song.

Which UVR model has the least vocal bleed?

There is no single answer that holds for every song. UVR users in the project's GitHub Discussions recommend different models and ensembles, and their advice changes as new models appear. Try two or three models on the same short passage of your song and compare the vocal gaps and the loud sections.

Does UVR's Ensemble mode remove bleed?

It can reduce it. Ensemble mode combines the outputs of several models, and users report fewer leftovers with some combinations. It also costs processing time and can make the voice duller, so compare the result with your best single model at matched loudness.

How do I clean a UVR vocal after separation?

Treat the gaps between phrases with a gate or region edits, then use a bleed remover for the bleed under the voice, and listen to what it removes. A plug-in such as de-leak does this inside your DAW, with a Residual monitor that plays only the removed signal.