Which UVR Model Should You Use for Instrumentals?

Model lists and leaderboards often show three scores: SDR, Fullness and Bleedless. All three grade how clean an instrumental is, each from a different angle:

• SDR: the overall score. How close the model's instrumental is to the correct answer.
• Bleedless: how little of the singer is left behind.
• Fullness: how much of the music survived, and whether it still sounds full.

No model ranks first on all three.¹ So before you pick one, decide which problem bothers you more: hearing a ghost of the original singer, or an instrumental that sounds thin.

How the scores are made

Scoring works a lot like checking an exam against an answer key.

The test songs come with the vocals and each instrument on separate tracks.² The tracks are mixed into a full song, the model splits it, and the instrumental it produces is compared with the original separate instrumental track. That original track is the answer key. The closer the model gets to it, the higher the score.

The scores in this article come from the public MVSEP leaderboard. Its test set has 100 songs in different styles, one minute each.²

SDR: one overall score

SDR stands for Signal-to-Distortion Ratio. Put simply, it compares two things: how loud the answer key is, and how loud the wrong parts of the result are.³ The smaller the wrong parts, the higher the SDR.

Wrong sound comes in three kinds: leftover vocals or other sounds that were not removed, noise, and odd sounds the model creates on its own, such as a bubbling, gurgling texture.⁴

SDR is measured in decibels (dB), but you can read it as a score. Because of how it is calculated, every 3 points higher means roughly half as much wrong sound, and every 10 points higher means one tenth as much.³

It is a strict score. On the same leaderboard, the answer key itself, saved as a common 128 kbps MP3, gets only 21.3 for the instrumental.¹ Even the original audio loses points after one ordinary compression. The best model in the table below scores 17.4.

But SDR is only a total. It does not tell you what went wrong, the same way an exam total does not tell you which questions you missed. Two models can both score 17 while one leaves more vocals and the other cuts part of the music, and they will sound very different. Researchers pointed out long ago that leftover sound, noise and odd artifacts need to be measured separately, because one number cannot describe them.⁴

Do not read much into gaps of a few tenths, either. One paper specifically criticizes using differences of around 0.1 to prove that one method is better.⁵

Bleedless: how much vocal is left

To bleed here means the vocal leaks into the instrumental. Bleedless means no leaking. The higher the score, the less of the singer you will hear.

Here is how it works: the result and the answer key are cut into many small frequency slices, and their loudness is compared slice by slice. Any sound the result has on top of the answer key counts as bleed.³

Fullness: was the music cut

Fullness looks at the opposite problem. The same slice-by-slice comparison is used, and any sound the result is missing compared with the answer key counts as removed by mistake.³ The higher the score, the more of the instrumental is kept.

Bleedless and Fullness are not a shared academic standard. They come from the code of an open-source model training project,³ and the MVSEP leaderboard can rank models by both.¹

Why the two scores pull against each other

Think about removing a person from a photo with an editing app.

Cut aggressively, and no trace of the person is left, but part of the background behind them often goes too. Cut carefully, and the background stays whole, but a faint outline of the person may remain.

Vocal removal works the same way. The more decisively a model removes, the less vocal is left and the higher its Bleedless, but the more likely it is to cut instruments too, which lowers its Fullness. The reverse is also true. Here are three instrumental results from the leaderboard. Higher is better for all three scores:¹

The first row has the best overall score and the least vocal left, but the lowest fullness. The last row keeps the most music and leaves the most vocal behind. The middle row sits between them on both.

Hear the difference

Below is the same 30-second chorus run through each of the three models in the table. All clips start at the same moment and none was turned up or down on its own. Switch tracks while it plays and the position stays where it is. The differences are small, so headphones help. Listen to the long held vocal notes for a faint trace of the singer, and to the drums and guitars for anything that sounds thinner.

Which score should you check

If you are making a karaoke or cover backing track, the thing you will notice first is a ghost of the original singer. Check Bleedless first. In the table, that points to the first row.

If you need the instrumental for sampling, remixing or practicing along, missing or thin instruments hurt more than a faint vocal ghost. Check Fullness first. That points to the last row. The middle row is a compromise between the two.

If you are not sure, start with the overall SDR score, and do not worry about differences of a few tenths.

Keep one more thing in mind. A leaderboard score is an average over 100 test songs, and your song may differ in style and recording quality. MVSEP itself notes that some test songs may have been seen by some models during training, a bit like seeing the exam questions in advance, which makes the comparison less fair.²

Let your ears decide

Scores help you narrow the list. The final call is a listening test. In the vocal remover from WaveReveal (formerly UVR Online), pick two or three models with different strengths and run the same song through each:

1. Listen to a quiet verse. Can you hear a faint trace of the original singer?
2. Listen to the busiest chorus. Do the drums, guitars or backing vocals sound thinner or disappear?
3. Use the result whose problem you mind least.

If you just want a usable instrumental fast, start with how to get an instrumental.