Music Source Separation on MUSDB18 HQ (test)
12.91SDR (Drums)BS-RoFormer
Evaluation Results
| Method | Links | |||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| BS-RoFormerTraining Data Configuration=extra data2026.03 | 12.91 | 11.99 | 13.32 | 9.01 | 12.72 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SCNet-largeExtra training data=235 songs2024.01 | 11.23 | 9.92 | 9.86 | 7.51 | 11.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SCNet-largeExtra training data=No2024.01 | 10.98 | 9.69 | 9.49 | 7.44 | 10.86 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Sparse HT DemucsExtra training songs=800, Fine-tuned=true2022.11 | 10.83 | 9.2 | 10.47 | 6.41 | 9.37 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| HT DemucsTraining Data Configuration=extra data2026.03 | 10.83 | 9.27 | 10.47 | 6.41 | 9.37 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SCNetExtra training data=235 songs2024.01 | 10.78 | 9.25 | 9.21 | 6.84 | 10.17 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SCNetExtra training data=No2024.01 | 10.51 | 9 | 8.82 | 6.76 | 9.89 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BSRNNExtra training data=1750 songs2024.01 | 10.15 | 8.97 | 8.16 | 7.08 | 10.47 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Band-Split RNNExtra training songs=1750, Only mixes used=true2022.11 | 10.15 | 8.97 | 8.16 | 7.08 | 10.47 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BSRNNTraining Data Configuration=extra data2026.03 | 10.15 | 8.97 | 8.16 | 7.08 | 10.47 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| HT DemucsExtra training data=800 songs2024.01 | 10.08 | 9 | 10.39 | 6.32 | 9.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| HT DemucsExtra training songs=800, Fine-tuned=true2022.11 | 10.08 | 9 | 10.39 | 6.32 | 9.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SIMO-BSRNNTraining Data Configuration=MUSDB18-HQ2026.03 | 10.06 | 8.54 | 7.8 | 6.56 | 9.73 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| HT DemucsExtra training songs=8002022.11 | 10.05 | 8.8 | 9.78 | 6.42 | 8.93 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| HT DemucsExtra training songs=1502022.11 | 9.51 | 8.49 | 9.76 | 6.13 | 8.56 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BS-RoFormerTraining Data Configuration=MUSDB18-HQ2026.03 | 9.49 | 9.8 | 11.31 | 7.73 | 10.66 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Hybrid DemucsExtra training data=800 songs2024.01 | 9.31 | 8.34 | 9.13 | 6.18 | 8.75 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Hybrid DemucsExtra training songs=8002022.11 | 9.31 | 8.34 | 9.13 | 6.18 | 8.75 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Hybrid DemucsTraining Data Configuration=extra data2026.03 | 9.31 | 8.34 | 9.13 | 6.18 | 8.75 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BSRNNExtra training data=No2024.01 | 9.01 | 8.24 | 7.22 | 6.7 | 10.01 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Band-Split RNNExtra training songs=02022.11 | 9.01 | 8.24 | 7.22 | 6.7 | 10.01 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BSRNNParams=37.6 M x 4, measurement_source=based on available code2023.09 | 9.01 | — | 7.22 | 6.7 | 10.01 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BSRNNTraining Data Configuration=MUSDB18-HQ2026.03 | 9.01 | 8.24 | 7.22 | 6.7 | 10.01 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TFC-TDF-UNet v3Speed=3.9x, overlap-add=true2023.06 | 8.96 | 8.04 | 7.53 | 6.32 | 9.34 | — | 9.59 | — | 8.45 | — | 8.44 | — | 6.86 | — | 8.34 | — | — | — | — | |
| BSRNNSpeed=0.7x2023.06 | 8.92 | 7.94 | 6.8 | 6.01 | 10.04 | — | 10.01 | — | 7.22 | — | 9.01 | — | 6.7 | — | 8.24 | — | — | — | — | |
| Hybrid DemucsSpeed=8.9x2023.06 | 8.87 | 7.53 | 7.76 | 5.39 | 8.11 | — | 8.13 | — | 8.76 | — | 8.24 | — | 5.59 | — | 7.68 | — | — | — | — | |
| TFC-TDF-UNet v3Speed=15.0x, overlap-add=false2023.06 | 8.81 | 7.9 | 7.36 | 6.19 | 9.22 | — | 9.38 | — | 8.28 | — | 8.01 | — | 6.77 | — | 8.11 | — | — | — | — | |
| IRM oracleExtra training songs=N/A2022.11 | 8.45 | 8.22 | 7.12 | 7.85 | 9.43 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TFC-TDF UNet v3Params=70.0 M2023.09 | 8.44 | — | 8.45 | 6.86 | 9.59 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TFC-TDF UNet v3Training Data Configuration=MUSDB18-HQ2026.03 | 8.44 | 8.34 | 8.45 | 6.86 | 9.59 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Hybrid Demucs (v3)Params=83.6 M x 4, bag of models=true2023.09 | 8.24 | — | 8.76 | 5.59 | 8.13 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Hybrid DemucsExtra training data=No2024.01 | 8.12 | 7.64 | 8.43 | 5.65 | 8.35 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Hybrid DemucsExtra training songs=02022.11 | 8.12 | 7.64 | 8.43 | 5.65 | 8.35 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Hybrid DemucsTraining Data Configuration=MUSDB18-HQ2026.03 | 8.12 | 7.64 | 8.43 | 5.65 | 8.35 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| HT DemucsExtra training data=No2024.01 | 7.94 | 7.52 | 8.48 | 5.72 | 7.93 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| HT DemucsExtra training songs=02022.11 | 7.94 | 7.52 | 8.48 | 5.72 | 7.93 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| HT DemucsParams=42.0 M2023.09 | 7.94 | — | 8.48 | 5.72 | 7.93 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| HT DemucsTraining Data Configuration=MUSDB18-HQ2026.03 | 7.94 | 7.52 | 8.48 | 5.72 | 7.93 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KUIELab-MDX-NetSpeed=8.5x2023.06 | 7.85 | 7.45 | 7.12 | 5.78 | 9.05 | — | 8.97 | — | 7.83 | — | 7.2 | — | 5.9 | — | 7.47 | — | — | — | — | |
| DTTNetParams=5.0 M x 42023.09 | 7.74 | — | 7.45 | 6.92 | 10.12 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DTTNetTraining Data Configuration=MUSDB18-HQ2026.03 | 7.74 | 8.06 | 7.45 | 6.92 | 10.12 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| D3NetExtra training data=1500 songs2024.01 | 7.36 | 6.68 | 6.2 | 5.37 | 7.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KUIELAB-MDX-NetExtra training songs=02022.11 | 7.33 | 7.54 | 7.86 | 5.95 | 9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KUIELab-MDX-NetParams=7.4 M x 42023.09 | 7.2 | — | 7.83 | 5.9 | 8.97 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KUIELab-MDX-NetTraining Data Configuration=MUSDB18-HQ2026.03 | 7.2 | 7.47 | 7.83 | 5.9 | 8.97 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TFC-TDF-UNet v2Speed=12.8x2023.06 | 6.87 | 7.03 | 6.85 | 5.44 | 8.96 | — | 9.05 | — | 7.61 | — | 6.4 | — | 5.7 | — | 7.19 | — | — | — | — | |
| ResUNetDecouple+Params=102.0 M x 42023.09 | 6.62 | — | 6.04 | 5.29 | 8.98 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DEMUCS2021.11 | 6.509 | 5.873 | 6.47 | 4.018 | 6.496 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CWS-PResUNet2023.06 | 6.38 | 6.77 | 5.93 | 5.84 | 8.92 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CWS-PResUNetParams=202.6 M x 42023.09 | 6.38 | — | 5.93 | 5.84 | 8.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CWS-PResUNetTraining Data Configuration=MUSDB18-HQ2026.03 | 6.38 | 6.77 | 5.93 | 5.84 | 8.92 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Our ApproachTraining Strategy=Transfer Learning with Jukebox, Fine-tuning=true, Initialization=Jukebox pre-trained weights2021.11 | 4.925 | 4.188 | 4.073 | 2.695 | 5.06 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Ground Truth2022.06 | 4.57 | 4.57 | 4.52 | 4.61 | 4.58 | — | — | — | — | — | — | — | — | — | — | 4.56 | — | — | — | |
| BigVGAN2022.06 | 4.41 | 4.26 | 4 | 4.25 | 4.37 | — | — | — | — | — | — | — | — | — | — | 4.26 | — | — | — | |
| BigVGAN-basefilter=included, snake_activation=included2022.06 | 4.39 | 4.2 | 4 | 4.14 | 4.36 | — | — | — | — | — | — | — | — | — | — | 4.11 | — | — | — | |
| HiFi-GAN (V1)2022.06 | 4.37 | 4.08 | 3.95 | 3.92 | 4.26 | — | — | — | — | — | — | — | — | — | — | 3.91 | — | — | — | |
| BigVGAN-basefilter=w/o filter, snake_activation=included2022.06 | 4.32 | 4.15 | 3.95 | 4.05 | 4.3 | — | — | — | — | — | — | — | — | — | — | 4.11 | — | — | — | |
| BigVGAN-basefilter=w/o filter, snake_activation=w/o snake2022.06 | 4.32 | 4.12 | 3.94 | 4.01 | 4.31 | — | — | — | — | — | — | — | — | — | — | 4.02 | — | — | — | |
| UnivNet-c32training_subset=train-clean-360, source=pretrained open-source repository2022.06 | 4.23 | 3.99 | 3.9 | 3.8 | 4.22 | — | — | — | — | — | — | — | — | — | — | 3.8 | — | — | — | |
| Wave-U-NetDomain=Waveform domain2021.11 | 4.22 | 3.23 | 3.21 | 2.25 | 3.25 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ScaledMixturePredictorContext=Music Demixing Challenge baseline2021.11 | 0.578 | 0.887 | 0.745 | 1.136 | 1.09 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BS-Locoformer (Medium)Encoder/Decoder=Band-split (62) [14], Params.=52.7 M2026.02 | — | — | — | — | — | 9.83 | 9.66 | 8.14 | 9.16 | 10.45 | 10.24 | 6.57 | 7.07 | 8.75 | 9.04 | — | — | — | — | |
| BS-Locoformer (Medium)ID=B1, Encoder/Decoder=Band-split (64) [15], Params.=55.5 M2026.02 | — | — | — | — | — | 9.59 | 9.44 | 8.9 | 10.5 | 10.43 | 10.76 | 6.24 | 7 | 8.79 | 9.42 | — | — | — | — | |
| BS-Locoformer (Medium)ID=B4, Encoder/Decoder=Band-split (64) [15], Params.=55.5 M, Training=Entire 100 songs, 8s duration, Note=Validation split not explicitly mentioned2026.02 | — | — | — | — | — | 10.02 | 9.77 | 9.28 | 10.67 | 10.94 | 11.02 | 6.81 | 7.56 | 9.26 | 9.75 | — | — | — | — | |
| BS-Locoformer (Small)ID=A1, Encoder/Decoder=Band-split (64) [15], Params.=34.7 M2026.02 | — | — | — | — | — | 9.01 | 8.88 | 8.21 | 9.31 | 9.93 | 10.27 | 5.87 | 6.43 | 8.26 | 8.72 | — | — | — | — | |
| BS-RoformerEncoder/Decoder=Band-split (62) [14], Params.=72.2 M, Note=Validation split not explicitly mentioned2026.02 | — | — | — | — | — | — | 10.66 | — | 11.31 | — | 9.49 | — | 7.73 | — | 9.8 | — | — | — | — | |
| BS-RoFormerType=D2026.04 | — | 2.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 2.88 | 3.41 | 2.47 | |
| BSRNNfinetuning=false2022.09 | — | — | — | — | — | 10.04 | 10.01 | 6.8 | 7.22 | 8.92 | 9.01 | 6.01 | 6.7 | 7.94 | 8.24 | — | — | — | — | |
| BSRNNfinetuning=true2022.09 | — | — | — | — | — | 10.47 | 10.47 | 7.2 | 8.16 | 9.66 | 10.15 | 6.33 | 7.08 | 8.42 | 8.97 | — | — | — | — | |
| BSRNNImplementation=Original Paper results2026.03 | — | — | — | — | — | 10.04 | 10.01 | 6.8 | 7.22 | 8.92 | 9.01 | 6.01 | 6.7 | 7.94 | 8.24 | — | — | — | — | |
| BSRNN (OLA)Implementation=Replicated, hop size=1.5 s2026.03 | — | — | — | — | — | 9.1 | — | 6.47 | — | 8.33 | — | 5.44 | — | 7.34 | — | — | — | — | — | |
| BSRNN (OLA)Implementation=Replicated, hop size=0.5 s2026.03 | — | — | — | — | — | 9.18 | — | 6.6 | — | 8.43 | — | 5.53 | — | 7.44 | — | — | — | — | — | |
| FaderImplementation=Replicated2026.03 | — | — | — | — | — | 9.16 | 9.14 | 6.51 | 7.72 | 8.55 | 8.07 | 5.44 | 5.68 | 7.42 | 7.65 | — | — | — | — | |
| generative MSS frameworkType=G2026.04 | — | 2.19 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 2.62 | 3.02 | 2.5 | |
| HTDemucs4Type=D2026.04 | — | 2.25 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 2.71 | 3.22 | 2.19 | |
| Hybrid Demucs2022.09 | — | — | — | — | — | — | 8.13 | 8.76 | — | 8.24 | — | 5.59 | — | 7.68 | — | — | — | — | — | |
| KUIELab-MDX-Net2022.09 | — | — | — | — | — | — | 8.97 | 7.83 | — | 7.2 | — | 5.9 | — | 7.47 | — | — | — | — | — | |
| oBSRNNModel Category=Optimized2026.03 | — | — | — | — | — | 9.78 | 9.81 | 8.38 | 9.85 | 10.28 | 10.31 | 5.85 | 6.31 | 8.57 | 9.07 | — | — | — | — | |
| oBSRNN-SIMOModel Category=Optimized2026.03 | — | — | — | — | — | 10.73 | 10.66 | 8.32 | 9.73 | 10.74 | 10.98 | 6.8 | 7.78 | 9.15 | 9.79 | — | — | — | — | |
| ResUNetDecouple+2022.09 | — | — | — | — | — | 8.98 | — | 6.04 | — | 6.62 | — | 5.29 | — | 6.73 | — | — | — | — | — | |
| SCNetType=D2026.04 | — | 2.17 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 2.65 | 2.89 | 2.33 | |
| SFC-Locoformer (Medium)ID=B2, Encoder/Decoder=SFC-Mamba (64), Params.=15.2 M2026.02 | — | — | — | — | — | 9.63 | 9.52 | 8.93 | 10.19 | 10.05 | 11.13 | 6.9 | 7.4 | 9.13 | 9.56 | — | — | — | — | |
| SFC-Locoformer (Medium)ID=B3, Encoder/Decoder=SFC-CA (64), Params.=16.0 M2026.02 | — | — | — | — | — | 10.18 | 9.94 | 9.15 | 10.85 | 11.08 | 11.25 | 7.12 | 7.77 | 9.38 | 9.95 | — | — | — | — | |
| SFC-Locoformer (Small)ID=A2, Encoder/Decoder=SFC-Mamba (64), Params.=5.1 M2026.02 | — | — | — | — | — | 9.29 | 8.97 | 8.29 | 9.25 | 10.51 | 10.36 | 6.45 | 6.97 | 8.63 | 8.86 | — | — | — | — | |
| SFC-Locoformer (Small)ID=A3, Encoder/Decoder=SFC-CA (64), Params.=5.8 M2026.02 | — | — | — | — | — | 9.61 | 9.39 | 8.7 | 9.8 | 10.75 | 10.72 | 6.71 | 7.17 | 8.95 | 9.27 | — | — | — | — | |
| SIMO stereo BSRNNEncoder/Decoder=Band-split (57) [16], Params.=107.7 M, Note=Validation split not explicitly mentioned2026.02 | — | — | — | — | — | 10.27 | 9.73 | 7.61 | 7.8 | 9.83 | 10.06 | 6.5 | 6.56 | 8.55 | 8.54 | — | — | — | — | |
| TS-BSMAMBA2Encoder/Decoder=Band-split (57) [16], Params.=35.5 M, Note=Uncommon validation split used2026.02 | — | — | — | — | — | 10.6 | 10.57 | 7.47 | 8.88 | 10.07 | 10.34 | 6.69 | 8.45 | 8.71 | 9.56 | — | — | — | — |