Music Source Separation on MUSDB18 (test)
7.12SDR (Bass)IRM oracle
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| IRM oracleDomain=N/A, # Param=N/A2019.09 | 7.12 | 8.45 | 7.85 | 9.43 | — | 8.21 | |
| TasNetExtra data=300 hours, Time domain operation=true2020.10 | 7.04 | 7.68 | 4.04 | 7.34 | 13.76 | 6.52 | |
| demucsExtra data=150 songs, Time domain operation=true2020.10 | 6.7 | 7.08 | 4.47 | 7.05 | — | 6.33 | |
| DemucsModel Category=Standard State-of-the-art, Training=w/. MUSDB18 Training2021.12 | 6.21 | 6.5 | 3.8 | 6.21 | — | — | |
| D3NetExtra data=1,500 songs / 93 hours, Time domain operation=false2020.10 | 6.2 | 7.36 | 5.37 | 7.8 | 14.26 | 6.68 | |
| Nachmani et. al.Domain=Time domain2020.10 | 5.88 | 6.15 | 4.32 | 6.92 | — | 5.82 | |
| NachmaniDomain=Time2020.10 | 5.88 | 6.15 | 4.32 | 6.92 | — | 5.82 | |
| DemucsDomain=Waveform, # Param=648.00M2019.09 | 5.83 | 6.08 | 4.12 | 6.29 | — | 5.58 | |
| demucsDomain=Time domain2020.10 | 5.83 | 6.08 | 4.12 | 6.29 | — | 5.58 | |
| Demucs2020.10 | 5.83 | 6.08 | 4.12 | 6.29 | — | 5.58 | |
| Conv-TasNet2020.10 | 5.66 | 6.08 | 4.37 | 6.81 | — | 5.73 | |
| GPOCM CUNet + LaSAFTDomain=Time-frequency, FFT window size=4096, Blocks=92020.10 | 5.63 | 5.68 | 4.87 | 7.33 | — | 5.88 | |
| Meta-TasNetTime-domain processing=true2020.02 | 5.58 | 5.91 | 4.19 | 6.4 | — | 5.52 | |
| Meta-TasNetDomain=Waveform, # Param=45.50M2019.09 | 5.58 | 5.91 | 4.19 | 6.4 | — | 5.52 | |
| Meta-TasNetDomain=Time domain2020.10 | 5.58 | 5.91 | 4.19 | 6.4 | — | 5.52 | |
| Meta-TasNetDomain=Time2020.10 | 5.58 | 5.91 | 4.19 | 6.4 | — | 5.52 | |
| Meta-TasNet2020.10 | 5.58 | 5.91 | 4.19 | 6.4 | — | 5.52 | |
| Meta-TasNetModel Category=Query based Model, Training=w/. MUSDB18 Training2021.12 | 5.58 | 5.91 | 4.19 | 6.4 | — | — | |
| SpleeterExtra data=24,097 songs / 79 hours, Time domain operation=false2020.10 | 5.51 | 6.71 | 4.02 | 6.68 | 12.54 | 5.78 | |
| X-UMX2020.10 | 5.43 | 6.47 | 4.64 | 6.61 | — | 5.79 | |
| TAK2 (MMDenseLSTM)Extra data=800 songs, Time domain operation=false2020.10 | 5.4 | 6.81 | 4.8 | 7.16 | 13.73 | 6.04 | |
| Sams-NetSlices=22019.09 | 5.32 | 6.59 | 4.04 | 6.44 | — | 5.6 | |
| IBM2018.05 | 5.3 | 6.87 | 6.42 | 7.5 | 10.83 | — | |
| Sams-NetSlices=42019.09 | 5.29 | 6.57 | 4.08 | 6.42 | — | 5.59 | |
| Sams-NetDomain=Spectrogram, # Param=3.70M, Slices=122019.09 | 5.25 | 6.63 | 4.09 | 6.61 | — | 5.65 | |
| D3NetExtra data=None, Time domain operation=false2020.10 | 5.25 | 7.01 | 4.53 | 7.24 | 13.52 | 6.01 | |
| D3Net (proposed)Dilation Type=Multidilated, Domain=Frequency domain2020.10 | 5.25 | 7.01 | 4.53 | 7.24 | 13.52 | 6.01 | |
| D3NetDomain=Time-frequency, FFT window size=40962020.10 | 5.25 | 7.01 | 4.53 | 7.24 | — | 6.01 | |
| Open UnmixTime-domain processing=false2020.02 | 5.23 | 5.73 | 4.02 | 6.32 | — | 5.36 | |
| Open-UnmixDomain=Spectrogram, # Param=8.90M2019.09 | 5.23 | 5.73 | 4.02 | 6.32 | — | 5.36 | |
| Sams-NetSlices=82019.09 | 5.23 | 6.56 | 4.06 | 6.61 | — | 5.62 | |
| UMXDomain=Frequency domain2020.10 | 5.23 | 5.73 | 4.02 | 6.32 | — | 5.33 | |
| Open UnmixModel Category=Standard State-of-the-art, Training=w/. MUSDB18 Training2021.12 | 5.23 | 5.73 | 4.02 | 6.32 | — | — | |
| Sams-NetSlices=12019.09 | 5.21 | 6.55 | 3.97 | 6.42 | — | 5.53 | |
| MMDenseNet#params [×10^6]=0.332018.05 | 5.19 | 6.27 | 4.64 | 3.87 | 15.41 | — | |
| MMDenseLSTM#params [×10^6]=1.222018.05 | 5.19 | 6.62 | 4.93 | 4.94 | 16.4 | — | |
| Sams-NetSlices=162019.09 | 5.19 | 6.45 | 4.01 | 6.51 | — | 5.54 | |
| D3Net standard dilationDilation Type=Standard, Domain=Frequency domain2020.10 | 5.19 | 6.61 | 4.53 | 7.12 | 13.39 | 5.86 | |
| MMDenseLSTMTime-domain processing=false2020.02 | 5.16 | 6.41 | 4.15 | 6.6 | — | 5.58 | |
| MMDenseLSTMDomain=Spectrogram, # Param=4.88M2019.09 | 5.16 | 6.41 | 4.15 | 6.6 | — | 5.58 | |
| TAK1 (MMDenseLSTM)Domain=Frequency domain2020.10 | 5.16 | 6.43 | 4.15 | 6.6 | 12.83 | 5.59 | |
| MMDenseLSTMModel Category=Standard State-of-the-art, Training=w/. MUSDB18 Training2021.12 | 5.16 | 6.41 | 4.15 | 6.6 | — | — | |
| UMX2020.10 | 5.07 | 6.04 | 4.28 | 6.25 | — | 5.41 | |
| UHL2 (BLSTM ensemble)Domain=Frequency domain2020.10 | 5.03 | 5.92 | 4.19 | 5.93 | 12.23 | 5.27 | |
| UHL2Model Category=Standard State-of-the-art, Training=w/. MUSDB18 Training2021.12 | 5.03 | 5.92 | 4.19 | 5.93 | — | — | |
| D3Net w/o dilationDilation Type=None, Domain=Frequency domain2020.10 | 4.97 | 6.37 | 4.21 | 6.86 | 13.19 | 5.6 | |
| Sams-NetSlices=182019.09 | 4.94 | 6.37 | 3.88 | 6.57 | — | 5.44 | |
| GRU dilation 1Domain=Frequency domain2020.10 | 4.86 | 5.86 | 4.65 | 6.85 | 13.4 | 5.58 | |
| DGRU-DConvDomain=Time-frequency, FFT window size=40962020.10 | 4.86 | 5.86 | 4.65 | 6.85 | — | 5.56 | |
| BLEND2#params [×10^6]=30.362018.05 | 4.72 | 6.25 | 4.75 | 4.33 | 16.04 | — | |
| Sams-NetSlices=242019.09 | 4.55 | 5.98 | 3.68 | 6.26 | — | 5.12 | |
| BLSTM#params [×10^6]=30.032018.05 | 3.99 | 5.28 | 4.06 | 3.43 | 14.51 | — | |
| Sams-Net (w/o Attention)Domain=Spectrogram, # Param=3.64M2019.09 | 3.89 | 4.71 | 3.22 | 4.8 | — | 4.16 | |
| 2048-d ST-SED-SEPModel Category=Zero shot Model, Training=w/o. MUSDB18 Training, embedding_dim=2048-d2021.12 | 3.8 | 5.44 | 3.05 | 6.15 | — | — | |
| Spect U-NetTime-domain processing=false2020.02 | 3.67 | 4.66 | 3.4 | 5.74 | — | 4.37 | |
| Spect U-NetDomain=Spectrogram, # Param=9.84M2019.09 | 3.67 | 4.66 | 3.4 | 5.74 | — | 4.37 | |
| Spec-U-NetModel Category=Standard State-of-the-art, Training=w/. MUSDB18 Training2021.12 | 3.67 | 4.66 | 3.4 | 5.74 | — | — | |
| 2048-d PANN-SEPModel Category=Zero shot Model, Training=w/o. MUSDB18 Training, embedding_dim=2048-d2021.12 | 3.38 | 5 | 2.86 | 6.06 | — | — | |
| Wave-U-NetTime-domain processing=true2020.02 | 3.21 | 4.22 | 2.25 | 3.25 | — | 3.23 | |
| Wave-U-NetDomain=Waveform, # Param=10.20M2019.09 | 3.21 | 4.22 | 2.25 | 3.25 | — | 3.23 | |
| WaveNetModel Category=Standard State-of-the-art, Training=w/. MUSDB18 Training2021.12 | 3.21 | 4.22 | 2.25 | 3.25 | — | — | |
| AQMSP-MeanModel Category=Query based Model, Training=w/. MUSDB18 Training2021.12 | 3.09 | 4.34 | 3.16 | 4.9 | — | — | |
| WKModel Category=Standard State-of-the-art, Training=w/. MUSDB18 Training2021.12 | 2.94 | 4 | 2.43 | 3.76 | — | — | |
| RGT1Model Category=Standard State-of-the-art, Training=w/. MUSDB18 Training2021.12 | 2.7 | 3.44 | 2.63 | 3.85 | — | — | |
| WaveNetTime-domain processing=true2020.02 | 2.49 | 4.13 | 0.41 | 3.35 | — | 2.6 | |
| WaveNetDomain=Waveform, # Param=3.30M2019.09 | 2.49 | 4.13 | 2.6 | 3.35 | — | 2.6 | |
| DeepConvSepTime-domain processing=false2020.02 | 0.17 | 3.14 | -2.13 | 2.37 | — | 0.89 | |
| DeepConvSepDomain=Spectrogram, # Param=0.32M2019.09 | 0.17 | 3.14 | -2.13 | 2.37 | — | 0.89 | |
| 527-d PANN-SEPModel Category=Zero shot Model, Training=w/o. MUSDB18 Training, embedding_dim=527-d2021.12 | -0.86 | 0.95 | -2.65 | 4.16 | — | — |