Skip to main content
Have a personal or library account? Click to login
Evaluation Platform and Benchmarks for Music Source Separation Cover
Read online

Figures & tables

Figure 1

Illustration of Bleedless and Fullness metrics. Blue represents what disappeared from the predicted stem compared to the original: the more blue there is, the worse the Fullness metric. Red represents what appeared as extra compared to the original stem signal (bleed from the other stems): the more red there is, the lower the Bleedless value.

Table 1

Selected models and their SDR scores for instrumental and vocal separation on the Multisong MVSep dataset.

Model and DescriptionSDR Instr.SDR Vocal
Sami‑bytedance‑v.1.1, private model18.1311.82
Ensemble: SCNet XL + BS‑Roformer + MelBand Roformer17.9211.62
BS‑Roformer (by MVSep)17.6211.31
MelBand Roformer (by Bas Curtiz)17.5511.24
MelBand Roformer (by Kimberley Jensen)17.3211.01
BS‑Roformer (by viperx) | Overlap 2 | dim_t 110117.2810.97
SCNet XL17.2710.96
SCNet Large17.0510.74
SCNet Small16.5610.25
VitLarge2316.099.78
ZFTurbo algorithm from MDX23 competition15.829.56
MDX‑Net (by Kimberley Jensen). Predict vocals part (FFT: 7680)15.709.36
MDX: UVR‑MDX‑NET_Main_427 (by UVR team). Predict vocals part (FFT: 6144)15.619.32
MDX: UVR‑MDX‑NET Inst HQ 2 (by UVR team). Predict instrumental part (FFT: 6144)15.229.11
Demucs4 HT algorithm trained on 800 songs, fine‑tuned for each stem14.738.42
Demucs3 algorithm trained on 800 songs14.548.24
Demucs4 HT algorithm trained on 800 songs, tuned for average SDR for 4 stems14.498.18
Demucs4 HT, 6 stems (additional piano and guitar)14.488.17
Demucs3 Model B, winning algorithm from MDX21 competition14.448.13
UVR vocal models, 0_HP2‑4BAND‑3090_4band_arch‑500m_113.597.28
Table 2

Selected models and their SDR metrics for bass, drums, and other separation on the Multisong MVSep dataset.

Model and DescriptionSDR BassSDR DrumsSDR Other
Ensemble: SCNet XL + MelBand Roformer + BS‑Roformer14.1413.578.10
SCNet XL13.8113.427.22
Sami‑bytedance‑v.1.1, private model13.7413.526.92
SCNet Large—13.01—
ZFTurbo algorithm from MDX23 competition12.5011.706.63
BS‑Roformer12.4912.82—
MelBand Roformer—12.53—
Demucs4 HT (fine‑tuned by MVSep)12.5212.04—
Demucs4 HT algorithm trained on 800 songs, fine‑tuned for each stem (shifts = 10, overlap = 0.95)12.2411.415.84
Demucs4 HT algorithm, fine‑tuned for each stem; (shifts = 1, overlap = 0.95) 10 times faster than the previous12.2411.405.84
Demucs4 HT algorithm trained on 800 songs, tuned for average SDR for 4 stems11.7410.905.57
Demucs4 HT, 6 stems (additional piano and guitar)11.4210.59—
Demucs3 algorithm trained on 800 songs11.2510.765.50
Demucs3 Model B, MDX21 competition winner10.6910.275.35
Table 3

SDR performance (dB) of selected models trained on MUSDB18‑HQ (100 songs) and evaluated on the MUSDB18 test set and the Multisong MVSep benchmark.

Model and DescriptionMUSDB18 TestMultisong MVSep
AvgBassDrumsVocalsOtherAvgBassDrumsVocalsOther
MDX23C17.155.777.939.235.687.028.407.737.364.57
SCNet Small (by starrytong)29.038.8910.449.906.898.8711.0710.798.275.34
SCNet Masked Small18.818.3210.4810.056.408.8611.0010.578.355.51
SCNet Large19.328.6310.8910.697.069.1911.1511.048.945.62
SCNet Large (by starrytong)29.709.3811.1510.947.319.2811.2711.239.055.57
SCNet XL19.809.2311.5111.057.419.7211.8711.499.326.19
SCNet XL IHF110.089.2311.8111.427.889.9211.9411.589.686.48
SCNet Masked XL IHF19.828.9111.6211.067.709.6711.5711.389.436.30
BS‑Roformer19.658.4811.6111.087.449.3811.0811.299.195.96
Table 4

Selected models and their SDR metrics on the Guitar MVSep dataset.

Model and DescriptionSDR
Sami‑bytedance‑v.1.1, private model8.27
Ensemble: BS‑Roformer + MelBand Roformer7.51
Ensemble: MDX23C + MelBand Roformer7.18
BS‑Roformer (by viperx)7.12
MelBand Roformer (by MVSep)7.02
MelBand Roformer (by viperx)6.85
MDX23C6.34
Demucs4 HT, 6 stems5.24
Table 5

Selected models and their performance metrics on the Piano MVSep dataset.

Model and DescriptionSDRL1FreqAuraMRSTFT
Sami‑bytedance‑v.1.1, private model7.3056.656.04
Ensemble: SCNet + MelBand Roformer6.2052.456.75
SCNet Large5.8951.566.66
MelBand Roformer (by viperx)5.7152.515.15
MDX23C5.5951.906.81
Demucs4 HT, 6 stems2.2341.213.69
Table 6

Selected models and their performance metrics on the Bowed Strings MVSep dataset.

Model and DescriptionSDRL1FreqAuraMRSTFT
Sami‑bytedance‑v.1.1, private model4.2544.625.32
MDX23C3.8442.796.18
UVR Online Strings model (based on MelBand Roformer)2.8639.984.12
Table 7

Selected models and their performance metrics on the Wind MVSep benchmark.

Model and DescriptionSDRL1FreqAuraMRSTFT
Sami‑bytedance‑v.1.1, private model7.5455.285.77
SCNet Large6.7651.956.31
MelBand Roformer (by MVSep)6.7352.955.32
X‑Minus Wind (MelBand Roformer)2.5141.783.62
Table 8

Selected models and their SDR metrics for drum components separation.

Model and DescriptionSDR
KickSnareTomshhCymbalshh‑Cymbals
SCNet XL, 5 stems (by MVSep)17.8912.5614.143.636.157.38
SCNet XL, 6 stems (by MVSep)17.7412.4314.243.395.917.19
SCNet XL, 4 stems (by MVSep)17.6112.3713.40——7.48
MelBand Roformer, 6 stems (by MVSep)17.4612.6413.695.057.067.91
MelBand Roformer, 4 stems (by MVSep)16.8512.4913.55——8.47
MDX23C, 5 stems (high fullness version)16.5011.4012.073.936.27—
MDX23C, 6 stems (by aufr33 and jarredou)14.549.7910.633.196.086.76
Demucs4 H, 4 stems (by inagoy)10.526.054.68——5.03
Table 9

Vocals and instrumental separation performance on the Multisong MVSep dataset, reporting Fullness, Bleedless, and SDR metrics. The Tune column indicates whether the model was tuned to favor the instrumental or vocal stem; ‘—’ denotes balanced models or unknown tuning.

Model NameArch.TuneVocalsInstrumental
FullnessBleedlessSDRFullnessBleedlessSDR
Kimberley Jensen1MelBand—16.6636.5111.0127.7146.7217.32
Bas Curtiz editionMelBand—16.3038.9411.1827.4947.0017.49
MVSep (ver. 2024.10)MelBand—16.9237.7811.2827.7147.2917.59
Unwa v1e2MelBandinstr.14.6726.8310.0638.8535.6816.37
Unwa big beta v5e3MelBandvoc.20.7832.0210.5925.6545.9016.90
Becruily instrum4MelBandinstr.15.7630.1510.1633.9340.5516.47
Becruily vocals5MelBandvoc.20.7231.2510.5528.2840.8516.86
Gabox v10 flowers6MelBandinstr.15.8525.3610.6637.6936.3416.97
Becruily deux7MelBand—22.8129.7811.3533.4142.1117.51
Unwa high instrum fullness8BS‑Roformerinstr.15.8524.2910.9434.7240.4317.25
Unwa HyperACE v2 instr9BS‑Roformerinstr.14.8133.0211.0937.9537.9717.40
Unwa HyperACE v2 vocals9BS‑Roformervoc.19.0734.1211.3927.3145.4817.70
MVSep (ver. 2025.07)BS‑Roformer—17.2338.2411.8927.8249.1218.20
MVSep SCNet XLSCNet—17.9626.9510.9628.7439.4217.27
MVSep SCNet XL (high fullness)SCNet—23.5025.3010.4034.0435.1516.60
MVSep SCNet XL IHFSCNet—17.9828.3111.1128.8740.3717.41
Becruily SCNet XL IHFSCNetinstr.22.7025.4810.8932.3138.1517.20
Table 10

Spearman correlation between objective metric differences and collective listener preference, by stem (N=30 pairs per stem).

Listening QuestionMetricVocals: ρ [95% CI]Instrumental: ρ [95% CI]
Which separation is better overall?ΔSDR+0.68 [+0.45, +0.80]−0.03 [−0.43, +0.37]
ΔL1Freq+0.74 [+0.54, +0.85]+0.09 [−0.32, +0.48]
ΔBleedless+0.51 [+0.20, +0.72]−0.13 [−0.44, +0.22]
ΔFullness+0.44 [+0.10, +0.69]+0.32 [−0.05, +0.64]
Which has fuller stem?ΔBleedless−0.01 [−0.34, +0.34]−0.63 [−0.75, −0.40]
ΔFullness+0.84 [+0.67, +0.92]+0.71 [+0.50, +0.85]
Which has less bleed from the other stem?ΔBleedless+0.50 [+0.18, +0.72]+0.44 [+0.08, +0.71]
ΔFullness−0.39 [−0.67, −0.02]−0.22 [−0.54, +0.15]
Which has fewer artifacts?ΔSDR+0.72 [+0.51, +0.84]+0.02 [−0.39, +0.40]
ΔL1Freq+0.71 [+0.50, +0.83]−0.06 [−0.41, +0.30]
ΔBleedless+0.50 [+0.14, +0.73]−0.16 [−0.50, +0.20]
ΔFullness−0.04 [−0.39, +0.33]+0.03 [−0.38, +0.49]
DOI: https://doi.org/10.5334/tismir.366 | Journal eISSN: 2514-3298
Language: English
Page range: 625 - 639
Submitted on: Jan 20, 2026
Accepted on: Sep 14, 2026
Published on: Nov 4, 2026
Published by: Ubiquity Press
In partnership with: Paradigm Publishing Services

© 2026 Roman Solovyev, Tatiana Habruseva, Jeremy Bonnot, Alexander Stempkovskiy, Ferdinand van Wijnen, published by Ubiquity Press
This work is licensed under the Creative Commons Attribution 4.0 License.