Bleedness and fullness leaderboard Python evaluation script by jarredou (prob. mirror), Torch version (with Bas Curtiz), used on Quality Checker

Page 36 of 65 · Edit this page in Google Docs ↗

Librosa version added to ZFTurbo training repo.

More detailed and reliable method of evaluation on multisong dataset than the popular SDR metric.

(old) Bas Curtiz’ evaluation chart with some Roformers tested with that method:
https://docs.google.com/spreadsheets/d/1pPEJpu4tZjTkjPh_F5YjtIyHq8v0SxLnBydfUBUNlbI/edit?usp=sharing (shortened version - it’s outdated - all metrics are rewritten in the models sections above)

For some newer models not on the list, you could search here for the model name, and bleedless/fullness metrics for new models are now provided in the evaluation description when you click on the result, but plenty of model evaluations have names not corresponding to final model names and were shared along with models on our Discord and later pasted to this document above.
Also, sorting by specific metric on MVSEP was added in June 2025, so you can track the exact evaluation by provided metrics in this document that way, or by searching Discord, but it can be difficult, as links to some older models’ evaluations were not indexed by the metrics bot, so once the evaluation was posted, the bot wasn’t showing metrics from the beginning, and not all models were evaluated along with the model release.

Explanations on the metric

Spectrogram difference showcase diagram

“Blue is what is missing from a separated stem (compared to a clean source).

Red is bleed in a separated stem.

White is perfect

(dB scale on the right seems wrong, I haven't checked, but it's not really important to see what is going on).

Same formula [can] be used for a metric, which would theoretically measure bleedness and fullness of the evaluated models

I think that for a metric, it's better to then separate negative values of diff array on one side, and keep positive values on other side, and average/scale each of them separately, so we get 2 scores, 1 for bleedness and 1 for fullness.

It has to be experimented further (and with better stft, it's only working on single chunk currently)

(Not sure that so high n_fft/mel_bins values are really needed, it was just nicer on the plot with that)”

“bleedless/fullness metrics are stft magnitude-only based and as they are discarding the phase data, they have some kind of blind spots.” - jarredou

Random noise added to results can increase fullness metric:

https://mvsep.com/quality_checker/entry/7709

https://mvsep.com/quality_checker/entry/7708

“l1 freq, the simplest way to explain it - it’s a mix between fullness and bleedless but without the noise issue (in a sense it’s the real fullness/bleedless metric) (...)
there’s no universal metric still sadly, we have to rely on a combination of them (and our ears)”

Q: How can you really spot when someone trains on the valid set anyway?

A: When it scores very well on valid tracks but is shit on anything else - becruily

Reservations and alternatives

“-l1_freq = bleedless (higher is cleaner)

-aura_mrstft = fullness (higher is fuller)

they maybe don’t have the issues fullness and bleedless have but I haven’t played to check that” - becruily

“aura_mrstft - is more perceptually relevant than SDR for musical content imo” - gilliaan

“Fullness and bleedless are not very reliable.

In my opinion it’s l1_freq and aura_mrstft both measure fullness and bleedless at once better (l1_freq has the edge because it also can show if model has phase issues like some fullness models)” - becruily

Read for discussion

“[The] problem with bleedless/fullness metric is that you can easily increase them by multiplying stem on constant.

Multiply predictions by 0.97 - it increases fullness and reduces bleedless

Multiply predictions by 1.03 - it greatly increases bleedless and reduces fullness” - ZFTurbo (metrics/discussion)

Q: I don't understand what fullness models do differently that result in this bleeding that is typical of fullness models

A: I would say that fullness models are leaving a larger ring of fire*/shell of the untargeted stem (link)

(I guess that with a different n_fft resolution or number of bands, that bleed noise would sound also differently with roformers iirc) - jarredou

Other metrics:

Log WMSE - good “at least for drums or anything rich in low frequency content” - jarredou

"It is a relatively new time-domain metric over SDR and SI-SDR that is not overly sensitive to low frequencies like SDR and can accurately evaluate silent intervals.

In addition, time-domain metrics can be evaluated for both amplitude and phase." - Unwa

Metrics ignore phase, so probably phase fixer won't affect fullness/bleeedless metric.

More by unwa:

“https://github.com/ZFTurbo/Music-Source-Separation-Training/blob/main/utils/metrics.py

First, metrics can be categorized as either time-domain or time-frequency domain.

SDR, SiSDR, and log_wmse are time-domain metrics.

l1_freq, aura_stft, and aura_mrstft are all STFT-based metrics and belong to the time-frequency domain.

Time-domain metrics assess how accurately the waveform matches the original sound, while time-frequency metrics compare how closely the spectrogram resembles the original sound's spectrogram.

Time-domain metrics are biased toward low frequencies and are less affected by high frequencies.

Time-frequency domain metrics can evaluate both low and high frequencies equally when based on a linear-scale spectrogram.

When based on Mel spectrograms, they exhibit a bias closer to human hearing.

However, the weakness of time-frequency metrics is their lack of phase consideration, since all metrics here are based on amplitude or power spectrograms.

The log_wmse metric is unique in that it allows weighting tailored to human hearing.

Although it is a time-domain metric, it is special in that it also takes the frequency domain into account.

The difference between aura_stft and aura_mrstft is that stft compares spectrograms of a single resolution, while mrstft compares mel spectrograms of multiple resolutions.”

For evaluating specific instrument stems, interesting read:

https://arxiv.org/abs/2507.06917v2

_____________________________

Top metrics of publicly available Roformers for instrumentals available for download

(as for 18.06.25)

Instrumental models sorted by instrumental fullness metric:

INSTV6N (41.68)>INSTV5N (40.47)>inst_Fv4Noise (40.40)>Inst V1e (38.87)>Inst Fv3 (38.71)>INSTV7N (36.83).

While V1e+ (37.89) might be already muddy in some cases

Instrumental models sorted by instrumental bleedless metric:

Gabox inst_fv7b

Fullness: 27.07 (worse than most vocal Mel-Roformers later below)

Bleedless: 47.49

Inst_GaboxFv7z

Fullness: 29.38

Bleedless: 44.95

Unwa BS-Roformer-Inst-FNO

Fullness: 32.03

Bleedless: 42.87

Unwa v2

Fullness: 31.85

Bleedless: 41.73

Inst_gaboxBv3

Fullness: 32.13

Bleedless: 41.69

Inst_GaboxFv8 (its replaced v2 variant)

Fullness: 33.22

Bleedless: 40.71

Becruily inst

Fullness: 33.98

Bleedless: 40.49

Gabox instv7plus

Fullness: 29.83

Bleedless: 39.36

Unwa HyperACE

Fullness: 36.91

Bleedless: 38.77

Unwa v1

Fullness: 35.69

Bleedless: 37.59

Gabox fv3

Fullness: 38.71

Bleedless: 35.62

Unwa v1e

Fullness: 38.87

Bleedless: 35.59

Gabox fv5

Fullness: 39.40

Bleedless: 33.49

Vocal models/ensembles sorted by instrumental bleedless metric:
(more muddy; Gabox and Unwa’s Revive models not evaluated yet):

Descriptions of the public models

MVSep BS-Roformer (2025.07.20) - the 2 previous versions got replaced on the site by it

Inst. Fullness 27.83

Inst. Bleedless 49.12

MVSep Ensemble 11.50 (2024.12.20)

Inst. Fullness 27.17

Inst. Bleedless 47.94

MVSep Ensemble (4 stem) 11.93 (2025.06.30)

Inst. Fullness 28.70

Inst. Bleedless 47.68

MVSep MelBand Roformer (2024.10) - ZF’s FT of Bas’ model)

Inst. Fullness 27.73

Inst. Bleedless 47.48

BS-RoFormer SW 6 stem (MVSEP/Colab/undef13 splifft)

Inst. Fullness 27.45

Inst. Bleedless 47.41

(use inversion from vocals and not mixed stems for better instrumental metrics)

MDX23 Colab fork v2.5 by jarredou

Inst. Fullness 28.02

Inst. Bleedless 47.24

(more noticeable bleeding/noise than MVSep Ensemble above)

Beta 6X

voc_fv4

(Good if you need less vocal residues than typical instrumental Roformers (even less than Mel Kim, FT2 Bleedless, or Beta 6X - makidanyee).

MVSEP BS PolarFormer Vocals (2026.05)

Inst. Fullness 17.68

Inst. Bleedless 47.18

MelBand Roformer Kim

Inst. Fullness 27.44

Inst. Bleedless 46.56 (or 46.87; and increases with overlap 8 up to ~47.12)

Kim | FT2 Bleedless (by Unwa)

Inst. Fullness 27.78

Inst. Bleedless 46.31

Beta 5e (by unwa)

Inst. Fullness 27.63 (bigger metric than Kim)

Inst. Bleedless 45.90

Kim | FT 2 (by unwa)

Inst. Fullness 28.36

Inst. Bleedless 45.58

Kim | FT (by unwa)

Inst. Fullness 29.18

Inst. Bleedless 45.36

MVSEP BS Roformer (2025.06)

Inst. fullness: 17.30

Inst. bleedless: 37.83

(can be still a good choice in case of some crossbleeding, vocal chops, or residues of reverbs or BGV)

MVSEP Ensemble 11.93 (also contains 2025.06)

Inst. fullness: 17.73

Inst. bleedless: 36.30

___

Outperformed vocal models for instrumental bleedless
(still metrics for instrumental stem, so after inversion if not duality)

SYHFT V3 (by SYH99999)

Fullness 28.07

Bleedless 45.15

Duality v1 (by unwa)

Fullness 29.08

Bleedless 43.26

Duality v2 (by unwa)

Fullness 28.03

Bleedless 44.16

Mel Becruily vocal

Fullness 28.25

Bleedless 40.95

SYHFT V2.5 (by SYH99999)

Fullness 28.60

Bleedless 40.34

Big SYHFT V1 (by SYH99999)

Fullness 28.48

Bleedless 44.81

Unwa beta 4

Fullness 26.29

Bleedless 44.71

SYHFT V4 and V5 were never publicly released

___

bleedless+fullness/2=avg

experimental avg metric for vocals (favours bleedless metric)

Bas' Edition - 27.72

FT2 bleedless - 27,54 | 2,49

24.10 - 27.44 | 2,21

FT2 - 26.84 | 2,23

FT - 26.58

5e - 26.42 | 1,54

voc_gabox - 26.38

voc_fv2 - 26.36

voc_fv3 - 26.06

Becruily - 25.99

beta 4 - 25.93

FullnessVocalModel - 25.91

voc_fv4 - 25.02

Other ensembles for UVR5

Best newer ensembles on the list at the top of the doc. Older configurations follow after the listed hidden results below.

For reference, read MVSEP’s SDR evaluation chart (UVR ensembles will appear later in the chart).

Be aware that some of the results on the chart above at the top are not from UVR5 or use different methods and code to achieve better results and might be not public/still WiP, e.g. the following:

Hidden leaderboard results (all SDR results provided for instrumentals,
Discord links below are dead, but at least some can be found by the search on Discord and by verifying the opened link address which initial URL hasn’t changed):

- Bas’ unreleased fullband vocal model epoch 299 + voc_ft - SDR 16.32)

- this older viperx’ unreleased custom weights code (newer one is up already), besides, “instrumental vX” entries are his ones (it rather utilizes public models with his own non-public weighted inference, and he gatekeeps it for more than since MDX23 results were published).

BTW. ebright is probably the 2nd place in MDX23, at least the result appeared in similar time like ByteDance. 2nd place decided not to publish their work.

- 32-bit higher SDR result of original multisong dataset uploaded as output (opposed to the previous 16-bit currently on top). “Multisong dataset | Original stems | bass/drums/other joined” is not a model!

- Bytedance v.0.2 - inst. SDR 17.26, now it’s outperformed by v.0.3 and is 17.28, now called 1.0),

-"MSS" - is probably ByteDance 2.0, not multi source stable diffusion, as BD's test files which were published were starting with MSS name before, but the first doesn't necessarily contradict the latter, although they said to use novel arch - SDR 18.13, and probably another one by ByteDance - SDR 18.75, let's call it 2.1, but seeing inconsistent vocal result vs previous one here, we have some suspicions that the result was manipulated at least for vocals (or stems were given from different model).

- Ripple app/SAMI-Bytedance on the chart is 16.59, also input files weren't lossless.

- BS-Roformer results by viperx posted in Training

BTW. model_mel_band_roformer_ep_617_sdr_11.5882 is Bas Curtiz model trained purely on multisong dataset as an experiment, and won’t give good results outside multisong dataset.

mel_band_roformer_ep_125_sdr_11.2069 is Bas Curtiz fine-tune model trained from ZFTurbo checkpoint, and it was shared with him under condition it will remain non-public/MVSEP exclusive.

____

Some of these models in the download center are visible after using the VIP code.

Older the best ensembles for UVR by SDR :

(some newer/better ones than these located at the top of the doc)

For 28.07.23

Kim Vocal 2 + MDX23C_D1581 + Inst HQ3 + Voc FT | Avg/Avg

For 28.07.23 (#4563)

Kim Vocal 1 + Kim Vocal 2 + MDX23C_D1581 + Inst HQ3 + Voc FT + htdemucs_ft | Avg/Avg

For 27.07.23 (#4561)

Kim Vocal 1 + Kim Vocal 2 + Kim Inst + MDX23C_D1581 + Inst HQ3 + Voc FT + htdemucs_ft | Avg/Avg (beta UVR)

For 24.06.23 (#3842)

Kim Vocal 1 + 2 + Kim Inst + HQ3 + Voc FT + htdemucs_ft | Avg/Avg | Chunks: ON

(but for ensembles instead of single models it can score better with chunks disabled)

[Consider using MDX23C_D1581 vocal model above as well, if ensemble in this arch works correctly, if not, perform manual ensemble, not sure here)

As for the very big ensemble from older synth leaderboard (2023-04-30):
MDX-Net: 292, 496, 406, 427, Kim Vocal 1, Kim Inst + Demucs ft

Optionally, with later released models - voc_ft and Kim Vocal 2 -

It doesn't score too good SDR-wise on newer synth dataset, since it uses older models which have better counterparts already. Synth dataset is not used for evaluations for a long time.

For 13.06.23 (#3322)

Inst HQ2 + 427 + Inst Main + Kim Inst + Kim Vocal 1 + 2 + Demucs FT | Avg/Avg | Chunks Batch | Spectral inversion OFF

Most probably you can safely replace Inst HQ2 with HQ3 and 4 (better SDR) getting a slightly better SDR in ensemble (it’s just not tested in ensemble yet).

But be aware that “The moment you introduce Instrumental models, there will be a bit of residue in the vocal output.

However, the SDR scores higher.

I'd say go with Vocal models only, if you care about your vocal output.”

The same is vice versa for instrumentals.

- Older ensemble configurations or custom settings with lower SDR

(but might be useful for some specific songs or genres if further info is given)

From public models, the best SDR on 14.04.23:

For instrumentals

And

For vocals

As of 01.01.23 the best SDR for vocals/instrumentals has:

-UVR-MDX-NET INST MAIN + UVR-MDX-NET Inst 3` + `kim vocal model fine tuned (old)` + `Demucs: v4 | htdemucs_ft - Shifts: 2 - Ensemble Algorithm: Avg/Avg`, chunk margin: 44100 (better SDR compared to 22050), denoise output on (-||- off), spectral inversion off (-||- on)

- MDX-Net: Kim vocal model fine-tuned (old) + UVR-MDX-NET_Main_427 + Demucs: v4 | htdemucs_ft - Ensemble Algorithm: Avg/Avg, Volume Compensation: Auto

(it sets `1.035` - the best for Kim (old) model vs other options)

Shifts: 10 - Overlap: 0.25

- a bit worse ensemble settings than both ensemble settings above SDR-wise:

UVR-MDX-NET Inst 3 (464) and “UVR-MDX-NET_Main_438” vocal model (main) and htdemucs_ft - Ensemble Algorithm: Average/Average

- Also good combo (for instrumentals, vocals in half of the cases):

MDX-Net: UVR-MDX-NET Inst Main

VR Arc: 7_HP2_UVR

Demucs: v4 | htdemucs_ft

Max Spec/Max Spec

- UVR-MDX-NET Inst 3 as a main model and 7_HP2-UVR as a secondary with the scale set to 75%

(Anjok 21.12.22: Personally, I found that using [it] produces the cleanest instrumental."

“It means the final track will be 25% hp2 model and 75% inst 3 (similar to ensemble feature, but you have more control over how strong you want the secondary model to be)”

- MDX-NET inst3 model (464) with secondary model 9_HP2_UVR 71% (hendrysetiadi: seems to get the best results with e.g. disco songs).

- Inst Main + 427 + Net 1 (CyPha-SaRin: was a pretty good combo. One big model, one medium, one small, pretty decent results across the board. If a song going to have problematic parts, it's going to have regardless of what combo you picked, it seems.)

- kim vocal 1 + instr 3 + full 403 + inst HQ 1 + full 292 + instr main with MAX/MAX (hendrysetiadi: i think that's the best combination of ensemble that i found)

- For Rock/Metal - The MDX-Net/VR Architecture ensemble with the Noise Reduction set between 5-10 (depending on the track) and Aggression to 10.

- For Pop - The MDX-Net/VR Architecture ensemble with the Noise Reduction set between 0-4 and Aggression to 10. (Anjok, 13.05.22)

- Here is another ensemble that I have tried myself "VR Arc: 1_HP-UVR x MDX-Net: Kim Vocal 1 x MDX-Net: UVR-MDX-NET: Inst HQ 1 x MDX-Net: UVR-MDX-NET: Inst HQ 2" All with the average/average ensemble (Mikey/K-Pop Filters)

- Inst HQ 1 & Main 427 are best for India

-VR: 7_HP2-UVR, MDX: Kim vocal 1, Inst 3, Inst Main, Main, htdemucs_ft

Max/Max, main pair: vocals/instrumental

"Instrumentals sound so good using these settings also. I can’t believe this is possible. What an amazing software. Thank you to whoever made this." StepsFan

- I got an ensemble that works well for loud and crazy tracks (this instance it's dariacore lol) - by knock:

Models: Inst HQ 3, Main, Voc FT

Ensemble Algorithm: Avg/Avg

MDX-Net settings:

Vol Comp: Auto

Segment Size: 4096 (you can go up to 6144 if you want to wait longer, 4096 has seemed to be perfect for me)

Overlap: Default (which I believe is 0.5)

Shift Conversion Pitch: -6 (semitones)

Match Freq Cut-off: Off

Denoise Output: Yes

Spectral Inversion: No

Mateus Contini's methods

#1 (old)

-“TIP! For busy songs: I was testing some ensembles trying to get Instrumental Stems with less volume variation (muddy), preserving guitar solos, pads the most and I had great results doing the following, for anyone interested:

Ensemble (Demucs + 5_HP-Karaoke with Max for Instrumental stem) - The result will be the Instruments + Backing Vocals and this preserves most of the guitar solos, pads and things that MDX struggles.

Instrumental Stem Output > Demucs to remove the Backing Vocals from the track - This pass will remove the rest of the Vocals. In some cases will be some minor leftovers that you can clean later with other methods.

I find the results better than Demucs alone/ MDX models or other ensembles for what I'm looking for. I'm not evaluating noise, but fuller instrumental Stems, trying to preserve most of it and also the cost (time) to do it.

Since I'm not interested, for this case, in doing manual work song by song and just use these stems to sing over it, I find the results great.” - Mateus Contini

Q: Do you mean that you process Demucs 2 times? Once for ensemble with VR then the result was processed using Demucs again?

A: You can add other models with the ensemble, like Demucs, VR_5-Karaoke and HQ3 for an extra, before processing again with Demucs.

Also, this method is very good for leave good backing vocals into the instrumentals (only the ensemble result). I find extracting bv from the Vocal Stem to be less effective, giving you less material (comparing if you would join the bv with instrumentals later)

M.Contini Method #2 (newer)

Well, I tried to improve the results of the method I posted, so here it is, for **anyone interested in get fuller Instrumentals**, with a bit of bleed in some songs, wielding great results overall.

I'm doing this in the UVR-gui. The idea behind it is to scoop the vocals little by little, so the instrumentals is preserved the most. The proccess requires 3 extractions. Here are the Ensembles:

1. pass Ensemble: 5_HP-Karaoke-UVR + Inst HQ3 + htdemucs - Min/ Max

- If the song doesn't have BV, this will already give you good Instrumental Stem results. If you have Vocals bleeding into the Instr, continue to pass 2, but sometimes jumping straight to pass3 will produce better results.

- If the song have BV, this you keep a fuller **Instrumental Stem with BV** in it. If you want to keep the BV, but there is some Main Vocals bleeding through the Instr, continue to pass 2.

2. pass Ensemble: Kim Vocal 2 + Inst HQ3 + MDX Karaoke 2 - Min/Max

- This pass will try to preserve the BV in the Instrumental Stem while removing Main Vocal bleed. You can stop here if you want the **Instrumental Stem with BV**

3. pass Ensemble: Kim Inst + Inst HQ3 + htdemucs - Min/Max

- This pass will try to remove BV from the instrumental Stem and other Main Vocal Bleed while keep the Instrumental fuller.

The idea behind it, is to have less volume variation where the vocals are extracted, leaving the Instrumental Stem less muddy. Since the extraction of the vocals is done little by little using the Min/Max, the Models will not be so aggressive. This is a great starting point if you want to improve further in a DAW or just sing over it. The Con is that, sometimes, the track will have tiny bleeds. If you try this method, please post the results here.

- -try this ensemble: 9_HP (10 agression) + HQ3 (chunks on) + demucs_ft, Min/Max

- it preserves most of the instruments.

M. Contini method #4 (new)

Another Ensemble suggestion for good instrumentals with minimized bleeding vocals and a bit of noise in some cases:

Ensemble: 9_HP + HQ3 + Demucs_6s (secondary model 50%: full_292) - Algorithm [min/max]

Configs:

9_HP Window[512], Agress[10], TTA[on], Post[off], High-End [off])

HQ3 Chunks[on] [auto], Denoise[on], Spectral[off]

Demucs_6s Chunks[on] [auto], Split[off], Combine[off], Spectral[off], Mixer[off], Secondary Model - Vocals/Instr [MDX-Inst_full_292] [50%]

Why Demucs_6s and not _ft - I compare them in some songs and 6s have less vocal bleed in the instrumental track.

Description:

The idea is to take the good bits of the models using only one from each Group (VR, MDX and Demucs). The secondary model on Demucs is to minimize some vocal bleeding with sustained notes that was happening in some songs.

Comparing the results from multiple models, I find that Chunks enabled on MDX and Demucs removes some bleeding vocals from the Instrumental track and gives better results overall. This ensemble in my machine completes in about 5 min per song (GTX 1070 8GB, 16GB RAM, Ryzen 1600x). [chunks have been replaced by newer method in newer UVR GUI versions]

____________

- “The best combo is the HQ instrument models ensemble average/average including HQ3/Main/Main Inst/Kim1/2/Kim Inst/demucs3 (mdx_extra)/htdemucs_ft/hdtdemucs6s” (MohammedMehdiTBER)

"Wow, I tried out the ensemble with all those models you said, and it actually sounds pretty good. There's a definitely more vocal bleed but in a saturated/detailed distortion type of way. I can't tell which one I like better, the ensemble sounds more full and has more detailed frequencies, but the vocal bleed is a lot more obvious. The HQ_3 by itself has almost no vocal bleed but sounds more thin and watery."

- Kim instr + mdx net instr3 + HQ2 + HQ3 + voc ft max/max

The result is so amazing… Now can hear more detail on instrumental result where before I cannot hear a bit of music parts. (Henry)

- "I am very much enjoying making an ensemble of HQ3 and MDX23C_D1581, then inverting the vocals into the instrumental and running that through hq3 with 0.5 overlap" (Rosé)

__________________________________

Ensembles for specific genres

Evaluation based on public models available at 23.04.23 and multisong dataset on MVSEP. The list might be outdated, as it doesn’t take all the current models into account.

SDR sorted by genre

By Bas Curtiz

"If we remove **Kim vocal 2**, so only those that are available right now will be taken into account:

- Ensemble Rating 1 scores highest on average overall

[Probably this one:

Kim vocal 2 + Kim FT other + Inst Main + 406 + 427 + htdemucs_ft | Avg/Avg

At least it was the best for the given date.

But now we have ensembles which score better.]

- Kim vocal 1 is best for Rock

- Kim vocal 1 & Ensemble Rating 1 are best for RnB/Latin/Soul/Funk

- MDX'23 Best Model is best for Pop

- Main 427 & MDX'23 Best Model are best for Other

- Main 427 & MDX'23 Best Model are best for Blues/Country

- Main 427 & Ensemble Rating 1 are best for Jazz

- Main 427 & Ensemble Rating 1 are best for Acoustic genres

- Ensemble Rating 1 is best for Beats

- Ensemble Rating 1 is best for Hip Hop

- Ensemble Rating 1 is best for House

Sheet where **Kim vocal 2 **is removed:

https://docs.google.com/spreadsheets/d/1ceXA7XKmECwnsQvs7a0S81XZOUokIXUN8ndsUDcYRcc/edit?usp=sharing"

Further single MDX-UVR models descriptions

E.g. used for ensembles above, but if a model has a cutoff, using ensemble with models/AIs without cutoff like Demucs 2-4 will fill the gap above. But it's still a good alternative for people without decent Nvidia GPUs or are force to use Colab.

UVR-MDX models naming scheme

All models called "main" are vocal models.

All models called "inst" and "inst main" are instrumentals.

NET-X [9.X/9.XXX in Colab] are vocal models

Kim vocal 1/2 (self-explanatory)

Inst main is 496

Kim other ft is Kim inst

Model labelled as just ‘main’ is vocal, and was reported to have the same checksums as 427 and 423, but it doesn't seem to be true as 427 and main have different SDR (427 has better SDR than main, so apparently main is 423 [CRC32: E3C998A6]).

- MDX HQ_1/2 models - excellent, vivid snares, no cutoff (22kHz) high quality, rarely worse results than narrowband inst1-3 models, HQ_2 might have slightly less loud snares, but can have fewer problems with removing some vocals from instrumentals

- MDX-UVR Inst 3 model (464) - 17.7 cutoff (the same cutoff as for Inst 1, 2 inst main, but maybe not applicable for vocals after inversion in Colab), it was the third-best single model in our SDR chart at the time, available in Colab update and UVR5 GUI with VIP models package - now available for free.

- Forth-best single model for instrumentals back then was inst main (496, MDX 2.1), then inst 1 and inst2.

- There was some confusion about MDX 2.1 model (iirc on x-minus) being vocal 438 (even 411), but it’s currently inst main.

- Full band MDX-Net models without cutoff (better SDR than Demucs 4 ft)

As for SDR, the epochs score is following: 292<403<386<(inst 1)<338<382<309<337

<450 (first final, HQ_1)<498 (HQ_2)<(inst3)<(Kim inst)<HQ_3<HQ_4

Epochs 292, 403 and 450 and newer are also in Colab (and in UVR5, older when VIP code is redeemed)

- (currently the best, maybe not single model, but custom ensemble, as for vocals) MDX23 in MVSEP beta,

and in UVR5 - Kim vocal model -

It's a further trained MDX-UVR vocal model from their last epoch (probably UVR-MDX-NET Main). It's based on a higher n_fft scale which uses more resources.

Not always gives that good results for instrumental as SDR may suggest, and also more people shares that opinion [both Colab and UVR users, so i’ts not due to no cutoff in Colab]).

In UVR5 generally for the best vocal result use vocal models, and for the best instrumental result use instrumental models or eventually 4 stem Demucs 4 ft.

"[Kim_Vocal_1] is an older model (November), than Kim uploaded at 2022-12-04 to" https://mvsep.com/quality_checker/leaderboard.php?sort=insrum&ensemble=0

(steps below no longer necessary, the model is added to GUI and these are the same models)

You can download her (so-called “old”) model from here (it still gets better results for vocals than inst 3 and main): https://drive.google.com/drive/folders/1exdP1CkpYHUuKsaz-gApS-0O1EtB0S82?usp=sharing

When you copy/paste the model in `C:\Users\YOURUSERNAME\AppData\Local\Programs\Ultimate Vocal Remover\models\MDX_Net_Models` It asks you to configure, hit Yes.

Then change `n_fft to 7680`."

For instrumentals, it gets worse results, frequently with more bleeding, and UVR manually applies cutoff above training frequency to instrumentals after inversion, to avoid some noise and possibly bleeding. Colab version of Kim model doesn’t have that cutoff, so instrumentals as a result of inversion have max 22kHz frequency (but UVR applies it to prevent some noise).

- (generally outperformed by models above) MDX-UVR 9.7 vocal model a.k.a. UVR-MDX-NET 1 (instrumental is done by inversion, older model) - available in Google Colab/mvsep (here 24 bit for instrumentals)/UVR5 GUI.

Compared to 9.682 NET 2 model, it might have better results on vocals, where 9.682 NET might have better results for instrumentals, but everything might still depend on a song. Generally, 9.7 model got better SDR both in Sony Demixing Challenge and on MVSEP. Generally, 438 vocal, or 464 inst_3 should give better results for instrumentals. 427 vocal model tends to give worse results for instrumentals than even this older 9.7/NET1 model.

More about MDX-UVR models -

If they don't have more vocal bleeding than GSEP, they’re better in filtering more vocal leftovers which sometimes GSEP tend to leave (scratches, additional vocal sounding sounds, also so-called “cuts” [short multiple lo-fi vocal parts] which GSEP doesn’t catch, but MDX-UVR does probably due to bigger dataset). But using single instrumental MDX-UVR models instead of ensemble will result in cut off of a training frequency (e.g. 17.7kHz or lower).

Also, MDX-UVR like GSEP may not have this weird constant "fuzz" which VR models tend to leave as vocal leftovers (but in other cases, 9.7 model can leave very audible vocal residues, so test out everything on this list, till you get the best result).

The 9.7 model (or currently newer models) is also good for cleaning inverts (e.g. when having lossy a cappella and regular song).

If you tested all the alternatives, and you stick to the MDX-UVR 9.7 for some song, and it doesn't have (too much) bleeding, to fine-tune the results you can try out two 9.6 models to check whether it's better for you than 9.7 in this specific case (they're available at least in HV Colab and UVR5 GUI).

Newer MDX-UVR 423 vocal model usually provides more audible leftovers than 9.7 model.

To further experiment with MDX-UVR results, and you’re stuck with Colab, you can enable Demucs 2 model on Colab to "ensemble" it with MDX-UVR model (although metrics say it slightly decreases SDR, I like what it does in hi-end - it was suspected at some point the SDR decreasing problems may come out from enabling chunking).

________________

- Demucs 4 (htdemucs_ft) - no cutoff, it’s 4 stem, but you can perform mixdown without vocals in Audacity for instrumental - sometimes it may give you louder snare than in GSEP, but usually muffled shakers compared to GSEP. Also, it will give you more vocal residues than GSEP and MDX-UVR 464 (Inst 3). 6 stem models gives more vocal residues than 4 stem model (ft is the best one and also outperformed mdx_extra model [better than mdx_extra_q - quantized) but in some cases that might be worth to check old mdx_extra model as well (but

- (outperformed in many cases when used at least as a single models)

VR-architecture models (Colab, CLI or UVR5 GUI) sometimes provide cleaner and less muddy results for instrumentals than single narrowband models of MDX or even GSEP, only if they do not output too much vocal bleeding (which really happens for VR models frequently - especially for heavily processed vocals in contemporary music), but bleeding also depends on specific model:

- E.g. 500m_1 (9_HP2-UVR) and MSB2 (7_HP2-UVR) models are the most aggressive in filtering vocals among VR models, but other, less aggressive VR models may provide better sounding, less spoiled instrumentals (only if it is not paid for with worse vocal bleeding [BTW. I haven’t heard the newest 2022 VR model yet (available at least in UVR5 GUI, maybe for Patreons, not sure]).

All parameters and settings corresponding to specific models you’ll find in “VR architecture models settings” section.

- VR models-only ensemble settings - if your track doesn’t have too many problems with bleeding using VR-models above, to fine-tune the results achieved with VR, and to get rid of some mud, and e.g. get better sounding drums in the mix, I generally recommend VR-architecture models ensemble with settings I described in the linked section above.

I'd say it's pretty universal, though the most time/resource-consuming method.

Also, these ensemble settings from the UVR HV Colab seem to make decent job for extracting vocals in some cases when above solutions failed (e.g. claps leftovers).

Check also demucs_6s with 9 HP UVR and gsep in min-specs mode

Also, UVR5 GUI has rewritten MDX, so it can use their Demucs-UVR models from Demucs 3 (I think mvsep doesn't provide ensembling for any MDX models):

- (generally outperformed by MDX-UVR 4xx models) Demucs-UVR models - 1 and 2 models beside "bag" are worth trying out (mainly 1) on their own if the results achieved with above methods still provide too much bleeding - better results than e.g. bare MDX-UVR 9.7 or VR models or even GSEP in some specific cases (available on MVSEP and UVR5 GUI). They're Demucs 3, 2 stem better trained models by UVR team. No cutoff - 22kHz.

_______________________________

- As for extracting -

Karaoke / Backing Vocals

(more up-to date, but less descriptive list at the top)

check MDX-UVR Karokee 2 model (available on MVSEP, UVR 5 GUI)

TL;DR - "Usually MDX B Karaoke has really good lead vocals and UVR Karaoke has really good backing vox”

"There are 3 good karaoke models (the ones I'm referring to are on mvsep.com [they seem to be no longer available there]). "MDX B (Karaoke)" seems to be the best at getting lead vocals from karaoke while "karokee_4band_v2_sn" (UVR) and "HP_KAROKEE-MSB2-3BAND-3090" (UVR) seem to be best for backing vocals. I recommend using a mix of the 3 to get as many layers as possible, and then use Melodyne to extract layers as best as possible. Then combine the filter results and Melodyne and you should have smthn that sounds pretty good" karokee_4band_v2_sn model might be not compatible with Colab (check mvsep or UVR5 GUI)

- Demix Pro may do a better job in B.V. than models on x-minus.

Even than the new model on x-minus since 01.02.23, but might be worth trying out on some songs (the problem is probably bound to MDX architecture itself).

"MDX in its pure form is too aggressive and removes a lot of backing vocals. However, if we apply min_mag_k processing, the results become closer to Demix Pro"

- Medley Vox

(installation tutorial)

For separating different voices, including harmonies or backing vocals check out this vocal separator, the demos sound quite good and Cyrus model has pretty similar results.

It's for already separated or original acapellas. Sometimes it gives better results than BVE models. Output sample rate is 24kHz, but it can be easily upscaled by AudioSR well.

Org. repository

https://github.com/jeonchangbin49/medleyvox

Old info (dead link):

https://media.discordapp.net/attachments/900904142669754399/1050444866464784384/Screenshot_81.jpg

How to get vocals stems by using specific models:

Song -> vocal model -> Voc & Inst

Vocal model -> Karaoke model -> Lead_Voc & Backing_Voc

Lead_Voc + Inst = Lead_Inst

- How to get backing vocals using x-minus

https://x-minus.pro/page/bv-isolation?locale=en_US

“Method two is terrible and I do not recommend it” - Aufr33

-If you have x-minus subscription, you can use chain mode for Karaoke as it currently gives the best results

How it probably works under the hood?

"On sitting down and reading https://discord.com/channels/708579735583588363/900904142669754399/1071599186350440540

It's a multistep process where it mixes a little bit from MDX's split vocals and instruments.

Then passes that mixture through the UVR v2 karaoke/backing vocals model.

Then with those results, it inverts the separated lead vocal, and adds it to the instrumental result"

- As for 4 stem separation, check GSEP or Demucs 4 (now check better MDX23 Colab by jarredou)

(other stem is usually the best in GSEP, bass in Demucs 4, rest depends also on a song, and as for drums, if you further process them in DAW using plugins, then Demucs 4 is usually better as it's lossless and supports up to 32-bit float output).

Demucs 4 has also experimental 6 stem feature. Guitar (can give good results) and piano (it's bad and worse than GSEP).

- As for free electric guitar and piano stems, currently GSEP and MVSEP models are the best, but paid Audioshake provides better results than GSEP. Also in GSEP "when the guitar model works (and it grabs the electric), the remaining 'other' stem often is a great way to hear acoustic guitar layers that are otherwise hidden.". LALAL.AI also has piano model and is “drastically” better than Demucs.

- From paid solutions for separating drums' sections, there are FactorSynth, UnmixingStation, or free Drumsep (but rather use MDX23C model).

- As for specific sounds separation, check Zero Shot Audio.

______

Cutoffs examination with spectrograms for various models and AIs, available in UVR5 GUI, along with examined times needed for each model to process on CPU or GPU (1700x/1080 Ti) by Bas Curtiz (cutoffs examination not applicable for MDX Colab where there is none unlike in UVR [it's to prevent noise]):
https://docs.google.com/spreadsheets/d/1R_pOURv8z9GmVkCt-x1wwApgAnplM9SHiPO_ViHWl1Q/edit#gid=23473506

Spreadsheet of songs that use Vocals as a melody with snippets how they separate on various models/AIs

http://vocalisolationtesting.x10.mx/

___

In below sections you’ll find more details, links, Colabs, all tools/AIs listed, more information about specific models as alternatives to experiment further (mostly MDX-UVR instrumental and vocal models available in UVR5 GUI and https://x-minus.pro/ai and MVSEP). I also provide some technicalities/troubleshooting everywhere when necessary.

_________________________________________________________________________