DTTNet (MUSDB18 Weights)

Paper: https://arxiv.org/abs/2309.08684
Original repository: https://github.com/junyuchen-cjy/DTTNet-Pytorch

Metrics on Multisong Dataset (Num overlap: 4)

Instr vocals sdr: 8.7892 (Std: 3.0910)
Instr vocals bleedless: 22.4991 (Std: 5.7524)
Instr vocals fullness: 14.2901 (Std: 4.2082)
Instr vocals l1_freq: 32.4076 (Std: 10.5796)
Instr bass sdr: 8.4670 (Std: 3.6563)
Instr bass bleedless: 20.1901 (Std: 12.6365)
Instr bass fullness: 18.5676 (Std: 11.7615)
Instr bass l1_freq: 45.7439 (Std: 15.0347)
Instr drums sdr: 7.5790 (Std: 3.2534)
Instr drums bleedless: 19.8468 (Std: 5.8063)
Instr drums fullness: 11.9787 (Std: 4.7591)
Instr drums l1_freq: 28.5485 (Std: 10.9713)
Instr other sdr: 5.1253 (Std: 2.9940)
Instr other bleedless: 16.9402 (Std: 6.3507)
Instr other fullness: 15.0643 (Std: 4.2500)
Instr other l1_freq: 29.9692 (Std: 9.9267)

Metrics on MUSDB18 Dataset (Num overlap: 4)

Instr vocals sdr: 10.3124 (Std: 3.3140)
Instr vocals bleedless: 25.6473 (Std: 13.9282)
Instr vocals fullness: 19.8120 (Std: 6.2593)
Instr vocals l1_freq: 49.1044 (Std: 11.5595)
Instr bass sdr: 7.1779 (Std: 4.3300)
Instr bass bleedless: 23.0092 (Std: 17.2864)
Instr bass fullness: 18.2872 (Std: 12.8031)
Instr bass l1_freq: 55.2532 (Std: 15.6209)
Instr drums sdr: 7.6618 (Std: 2.9982)
Instr drums bleedless: 22.9381 (Std: 7.6466)
Instr drums fullness: 14.4343 (Std: 4.5705)
Instr drums l1_freq: 32.8488 (Std: 10.6579)
Instr other sdr: 6.2039 (Std: 2.5146)
Instr other bleedless: 21.2702 (Std: 7.7841)
Instr other fullness: 18.3491 (Std: 7.1196)
Instr other l1_freq: 35.3127 (Std: 10.8529)

MVSep Mega 53 Stems model

First version of model with 53 stems. It's based on BS Roformer architecture.

List of stems: ['accordion', 'acoustic-guitar', 'back-vocal', 'banjo', 'bass', 'bassoon', 'bells', 'bowed_strings', 'brass', 'cello', 'clarinet', 'congas', 'digital-piano', 'dobro', 'double-bass', 'drums', 'electric-guitar', 'flute', 'french-horn', 'glockenspiel', 'guitar', 'harmonica', 'harp', 'harpsichord', 'hh', 'keys', 'kick', 'lead-vocal', 'mandolin', 'marimba', 'oboe', 'organ', 'percussion', 'piano', 'saxophone', 'sitar', 'snare', 'strings', 'synth', 'tambourine', 'timpani', 'toms', 'triangle', 'trombone', 'trumpet', 'tuba', 'ukulele', 'viola', 'violin', 'vocal', 'wind', 'wind-chimes', 'woodwind']

There are some model notes:

• We are continuing to refine and improve the model.
• The model is memory-intensive, even with a batch size of 1. I recommend at least 16 GB of VRAM.
• Currently, the model’s performance on individual stems may be lower than the specialized models available on the MVSep site.
• Stems do not sum up to the original mixture, as some may contain overlapping information. For example, the "vocals" stem includes both lead and backing vocals.
• The model returns all stems by default, but filtering out empty stems during post-processing is a logical next step.

BS PolarFormer (Weights)

Version of BS Roformer based on PoPE (Polar Coordinate Positional Embeddings).

Arxiv paper: https://arxiv.org/pdf/2509.10534

BS Mamba2 (Weights)

Model trained only on MUSDB18 dataset (100 songs)

Multisong dataset metrics:

Num overlap: 2
Instr vocals sdr: 8.8272 (Std: 3.5375)
Instr vocals bleedless: 31.4882 (Std: 10.2950)
Instr vocals fullness: 13.1359 (Std: 4.1619)
Instr vocals l1_freq: 33.3577 (Std: 11.3776)

MUSDB18 Test Dataset metrics:

Num overlap: 2 
Instr vocals sdr: 10.8644 (Std: 3.7046)
Instr vocals bleedless: 39.1770 (Std: 20.2013)
Instr vocals fullness: 17.2136 (Std: 5.9952)
Instr vocals l1_freq: 51.1858 (Std: 12.4056)

BS Conformer (Medium)

MSUDB18 Test:

Num overlap: 2
Instr drums sdr: 10.9636 (Std: 3.5740)
Instr drums bleedless: 40.1846 (Std: 12.2989)
Instr drums fullness: 15.3477 (Std: 5.6870)
Instr drums l1_freq: 43.1116 (Std: 11.7196)
Instr bass sdr: 8.1117 (Std: 4.4536)
Instr bass bleedless: 40.1839 (Std: 24.3516)
Instr bass fullness: 13.1798 (Std: 11.4510)
Instr bass l1_freq: 57.7922 (Std: 15.7611)
Instr other sdr: 7.0391 (Std: 2.9666)
Instr other bleedless: 29.7821 (Std: 11.1954)
Instr other fullness: 17.8183 (Std: 6.0078)
Instr other l1_freq: 37.9554 (Std: 12.0237)
Instr vocals sdr: 10.6349 (Std: 3.6795)
Instr vocals bleedless: 38.0876 (Std: 17.5646)
Instr vocals fullness: 16.5019 (Std: 5.7524)
Instr vocals l1_freq: 50.4559 (Std: 11.5570)
Metric avg sdr        : 9.1873
Metric avg bleedless  : 37.0596
Metric avg fullness   : 15.7119
Metric avg l1_freq    : 47.3288

MusltiSong Dataset:

Num overlap: 2
Instr drums sdr: 10.7364 (Std: 4.4866)
Instr drums bleedless: 37.6237 (Std: 11.2495)
Instr drums fullness: 11.4404 (Std: 3.3226)
Instr drums l1_freq: 36.2296 (Std: 11.8242)
Instr bass sdr: 10.3603 (Std: 4.2584)
Instr bass bleedless: 42.4855 (Std: 26.5982)
Instr bass fullness: 13.3310 (Std: 9.9044)
Instr bass l1_freq: 50.5069 (Std: 15.4443)
Instr other sdr: 5.5397 (Std: 3.3859)
Instr other bleedless: 20.9410 (Std: 9.0741)
Instr other fullness: 14.0600 (Std: 4.1029)
Instr other l1_freq: 31.2465 (Std: 10.0779)
Instr vocals sdr: 8.7594 (Std: 3.3987)
Instr vocals bleedless: 33.0732 (Std: 11.9230)
Instr vocals fullness: 13.1176 (Std: 4.2827)
Instr vocals l1_freq: 33.4560 (Std: 10.8730)
Metric avg sdr        : 8.8489
Metric avg bleedless  : 33.5308
Metric avg fullness   : 12.9872
Metric avg l1_freq    : 37.8597

SCNet Masked XL IHF Weights

New version of SCNet (scnet_masked) with mask - allows to reduce noise. There are weights for large XL IHF version trained on MUSDB18.

MUSDB18 Test Metrics:

Num overlap: 4
Instr drums sdr: 11.6207 (Std: 3.3777)
Instr drums bleedless: 26.9811 (Std: 9.6603)
Instr drums fullness: 20.2270 (Std: 8.8575)
Instr drums l1_freq: 43.1023 (Std: 11.7691)
Instr bass sdr: 8.9183 (Std: 4.8270)
Instr bass bleedless: 27.1082 (Std: 17.1666)
Instr bass fullness: 16.3376 (Std: 10.3509)
Instr bass l1_freq: 58.6868 (Std: 15.5151)
Instr other sdr: 7.7079 (Std: 3.0485)
Instr other bleedless: 24.1793 (Std: 9.8236)
Instr other fullness: 20.2322 (Std: 6.8875)
Instr other l1_freq: 38.1973 (Std: 12.5904)
Instr vocals sdr: 11.0675 (Std: 3.7956)
Instr vocals bleedless: 26.4048 (Std: 11.8723)
Instr vocals fullness: 20.4435 (Std: 7.2946)
Instr vocals l1_freq: 50.6262 (Std: 11.3240)
Metric avg sdr        : 9.8286
Metric avg bleedless  : 26.1683
Metric avg fullness   : 19.3101
Metric avg l1_freq    : 47.6531

Multisong dataset Metrics:

Num overlap: 4
Instr drums sdr: 11.3847 (Std: 5.0454)
Instr drums bleedless: 23.6516 (Std: 8.3218)
Instr drums fullness: 15.8085 (Std: 5.4690)
Instr drums l1_freq: 36.6667 (Std: 12.2488)
Instr bass sdr: 11.5754 (Std: 4.9476)
Instr bass bleedless: 23.8556 (Std: 15.0766)
Instr bass fullness: 21.7412 (Std: 14.2511)
Instr bass l1_freq: 52.6289 (Std: 16.7418)
Instr other sdr: 6.3082 (Std: 3.4198)
Instr other bleedless: 18.6385 (Std: 7.7289)
Instr other fullness: 16.5832 (Std: 4.5989)
Instr other l1_freq: 32.4004 (Std: 10.3759)
Instr vocals sdr: 9.4303 (Std: 3.2363)
Instr vocals bleedless: 24.0554 (Std: 7.9791)
Instr vocals fullness: 16.1403 (Std: 4.4971)
Instr vocals l1_freq: 34.2907 (Std: 10.5988)
Metric avg sdr        : 9.6747
Metric avg bleedless  : 22.5503
Metric avg fullness   : 17.5683
Metric avg l1_freq    : 38.9967

SCNet Masked Small Weights

New version of SCNet (scnet_masked) with mask - allows to reduce noise. There are weights for Small version trained on MUSDB18.

MUSDB18 Test Metrics:

Num overlap: 4
Instr drums sdr: 10.4807 (Std: 3.1034)
Instr drums bleedless: 23.5738 (Std: 9.6661)
Instr drums fullness: 19.3672 (Std: 8.8171)
Instr drums l1_freq: 39.7583 (Std: 11.5116)
Instr bass sdr: 8.3271 (Std: 4.5202)
Instr bass bleedless: 28.8302 (Std: 17.6783)
Instr bass fullness: 15.2740 (Std: 10.2804)
Instr bass l1_freq: 57.3428 (Std: 14.7919)
Instr other sdr: 6.4003 (Std: 3.5504)
Instr other bleedless: 21.1916 (Std: 8.4212)
Instr other fullness: 18.6111 (Std: 6.4006)
Instr other l1_freq: 34.6862 (Std: 10.9688)
Instr vocals sdr: 10.0517 (Std: 3.3388)
Instr vocals bleedless: 26.1827 (Std: 13.2979)
Instr vocals fullness: 17.8171 (Std: 6.8853)
Instr vocals l1_freq: 47.6337 (Std: 11.5299)
Metric avg sdr        : 8.8149
Metric avg bleedless  : 24.9446
Metric avg fullness   : 17.7673
Metric avg l1_freq    : 44.8552

Multisong dataset Metrics:

Num overlap: 4
Instr drums sdr: 10.5715 (Std: 4.7403)
Instr drums bleedless: 21.5980 (Std: 7.4977)
Instr drums fullness: 15.0217 (Std: 4.9323)
Instr drums l1_freq: 34.3450 (Std: 11.8224)
Instr bass sdr: 11.0098 (Std: 4.6274)
Instr bass bleedless: 24.4461 (Std: 16.1432)
Instr bass fullness: 21.4828 (Std: 15.0182)
Instr bass l1_freq: 51.1773 (Std: 16.3137)
Instr other sdr: 5.5176 (Std: 3.0846)
Instr other bleedless: 16.5156 (Std: 5.8950)
Instr other fullness: 16.1805 (Std: 4.3133)
Instr other l1_freq: 30.2492 (Std: 9.6913)
Instr vocals sdr: 8.3510 (Std: 3.0380)
Instr vocals bleedless: 21.5931 (Std: 7.6635)
Instr vocals fullness: 14.0202 (Std: 4.2627)
Instr vocals l1_freq: 31.0896 (Std: 10.2234)
Metric avg sdr        : 8.8624
Metric avg bleedless  : 21.0382
Metric avg fullness   : 16.6763
Metric avg l1_freq    : 36.7153

SCNet XL IHF Weights

Weights were obrained with training only on MUSDB18 training data. Main changes in config is related to high frequencies, so model should give better results in this part of spectrogram.

Results on MUSDB18 test data:

Instr drums sdr: 11.8155 (Std: 3.2582)
Instr drums bleedless: 28.3940 (Std: 11.3584)
Instr drums fullness: 20.3435 (Std: 7.6190)
Instr drums l1_freq: 43.2681 (Std: 11.5372)
Instr bass sdr: 9.2360 (Std: 5.0752)
Instr bass bleedless: 24.9027 (Std: 13.8338)
Instr bass fullness: 17.0699 (Std: 11.1160)
Instr bass l1_freq: 58.6811 (Std: 15.1790)
Instr other sdr: 7.8828 (Std: 3.1603)
Instr other bleedless: 24.7929 (Std: 9.5862)
Instr other fullness: 19.6569 (Std: 6.9291)
Instr other l1_freq: 38.2785 (Std: 12.4884)
Instr vocals sdr: 11.4219 (Std: 3.9988)
Instr vocals bleedless: 23.4196 (Std: 9.7727)
Instr vocals fullness: 20.3662 (Std: 7.8029)
Instr vocals l1_freq: 50.3840 (Std: 11.0768)
Metric avg sdr        : 10.0891
Metric avg bleedless  : 25.3773
Metric avg fullness   : 19.3591
Metric avg l1_freq    : 47.6529

Results on Multisong dataset:

Instr drums sdr: 11.5838 (Std: 5.3371)
Instr drums bleedless: 24.4482 (Std: 8.1998)
Instr drums fullness: 16.0051 (Std: 4.1486)
Instr drums l1_freq: 37.1487 (Std: 12.4679)
Instr bass sdr: 11.9436 (Std: 5.2532)
Instr bass bleedless: 23.0196 (Std: 11.7992)
Instr bass fullness: 20.1038 (Std: 13.4688)
Instr bass l1_freq: 52.8341 (Std: 16.9122)
Instr other sdr: 6.4873 (Std: 3.6878)
Instr other bleedless: 19.0051 (Std: 8.0195)
Instr other fullness: 16.3094 (Std: 5.0124)
Instr other l1_freq: 32.6761 (Std: 10.9044)
Instr vocals sdr: 9.6896 (Std: 3.5071)
Instr vocals bleedless: 22.7375 (Std: 7.0052)
Instr vocals fullness: 15.9332 (Std: 4.2005)
Instr vocals l1_freq: 34.3502 (Std: 10.8328)
Metric avg sdr        : 9.9261
Metric avg bleedless  : 22.3026
Metric avg fullness   : 17.0878
Metric avg l1_freq    : 39.2523

SCNet Tran Weights

Pretrained weights for new SCNet Tran model. It's model where LSTM blocks replaced with transformer blocks. Weights were obtained on MUSDB18 dataset only.

Multisong dataset metrics

Num overlap: 2
Instr drums sdr: 10.8707 (Std: 4.9156)
Instr drums bleedless: 21.1680 (Std: 6.7914)
Instr drums fullness: 14.7370 (Std: 4.3909)
Instr drums l1_freq: 34.3470 (Std: 11.9350)
Instr bass sdr: 10.9921 (Std: 4.7437)
Instr bass bleedless: 19.7168 (Std: 9.4653)
Instr bass fullness: 18.9573 (Std: 13.3901)
Instr bass l1_freq: 50.0550 (Std: 15.6142)
Instr other sdr: 5.6386 (Std: 3.1793)
Instr other bleedless: 15.4912 (Std: 5.5533)
Instr other fullness: 16.9250 (Std: 4.4698)
Instr other l1_freq: 30.2603 (Std: 9.8843)
Instr vocals sdr: 8.4272 (Std: 3.2268)
Instr vocals bleedless: 18.7054 (Std: 5.8792)
Instr vocals fullness: 15.2973 (Std: 4.1997)
Instr vocals l1_freq: 31.0975 (Std: 10.2741)
Metric avg sdr        : 8.9821
Metric avg bleedless  : 18.7704
Metric avg fullness   : 16.4791
Metric avg l1_freq    : 36.4400

MUSDB18 dataset metrics

Num overlap: 2
Instr drums sdr: 10.8139 (Std: 2.9125)
Instr drums l1_freq: 39.1727 (Std: 10.2864)
Instr drums fullness: 17.6692 (Std: 6.2694)
Instr drums bleedless: 25.0680 (Std: 10.1685)
Instr bass sdr: 8.0731 (Std: 4.6414)
Instr bass l1_freq: 55.2888 (Std: 13.7685)
Instr bass fullness: 14.9235 (Std: 9.6630)
Instr bass bleedless: 20.8816 (Std: 10.5097)
Instr other sdr: 6.8419 (Std: 2.8909)
Instr other l1_freq: 35.2425 (Std: 11.2693)
Instr other fullness: 19.2561 (Std: 6.6855)
Instr other bleedless: 20.1550 (Std: 7.3377)
Instr vocals sdr: 9.9799 (Std: 3.5204)
Instr vocals l1_freq: 45.7778 (Std: 11.0381)
Instr vocals fullness: 20.0345 (Std: 7.0781)
Instr vocals bleedless: 17.8214 (Std: 7.0943)
Metric avg sdr        : 8.9272
Metric avg l1_freq    : 43.8704
Metric avg fullness   : 17.9708
Metric avg bleedless  : 20.9815

SCNet XL

Pre-trained weights on MUSDB18HQ. Metrics table