Speech Separation on WSJ0-2Mix
25.1SI-SNRi (dB)SepReformer-L
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SepReformer-LParams (M)=59.4, MACs (G/s)=155.5, Dynamic Mixing=true2024.06 | 25.1 | 25.2 | |
| ISCIT+Params (M)=58.4, MACs (G/s)=252.2, Dynamic Mixing=true, Additional speaker information=true2024.06 | 24.3 | 24.4 | |
| Mossformer2(L)Params (M)=55.7, Dynamic Mixing=true2024.06 | 24.1 | — | |
| SFSRNetParams (M)=59.0, MACs (G/s)=466.2, Dynamic Mixing=true2024.06 | 24 | 24.1 | |
| Separate And DiffuseDynamic Mixing=true2024.06 | 23.9 | — | |
| QDPNParams (M)=200.0, Dynamic Mixing=true2024.06 | 23.6 | — | |
| QDPN+DMModel size=200M2022.10 | 23.6 | — | |
| Mossformer(L)Params (M)=42.1, MACs (G/s)=86.1, Dynamic Mixing=true2024.06 | 22.8 | — | |
| SepformerParams (M)=26.0, MACs (G/s)=86.9, Dynamic Mixing=true2024.06 | 22.3 | 22.5 | |
| SepFormerModel size=26M, Domain=Learnable latent domain-based, Data Augmentation=true2021.10 | 22.3 | 22.4 | |
| SepFormer+DMModel size=26M, GMACs=69.62022.10 | 22.3 | 22.4 | |
| WavesplitModel size=29M, Domain=Learnable latent domain-based, Data Augmentation=true2021.10 | 22.2 | 22.3 | |
| QDPNParams (M)=2002023.09 | 22.1 | — | |
| SPGMParams (M)=26.22023.09 | 22.1 | — | |
| SFSRNetParams (M)=59.02023.09 | 22 | — | |
| SepformerParams (M)=25.72023.09 | 21.6 | — | |
| DPTNET-SRSSNModel size=5.7M, Domain=Learnable latent domain-based2021.10 | 21.2 | 21.4 | |
| WaveSplitParams (M)=29.0, Dynamic Mixing=true2024.06 | 21 | 21.2 | |
| WavesplitModel size=29M, Domain=Learnable latent domain-based, Data Augmentation=false2021.10 | 21 | 21.2 | |
| WavesplitParams (M)=292023.09 | 21 | — | |
| DPRNN-SRSSNModel size=7.5M, Domain=Learnable latent domain-based2021.10 | 20.5 | 20.7 | |
| SepFormerModel size=26M, Domain=Learnable latent domain-based, Data Augmentation=false2021.10 | 20.4 | 20.5 | |
| DPTNET-TASNETModel size=2.7M, Domain=Learnable latent domain-based2021.10 | 20.2 | 20.6 | |
| DPTNetParams (M)=2.72023.09 | 20.2 | — | |
| MULCAT#params=7.5M2020.02 | 20.1 | — | |
| Nachmani et al.Model size=7.5M, Domain=Learnable latent domain-based2021.10 | 20.1 | — | |
| PapezParams (M)=1.472023.09 | 19.2 | — | |
| SuDORM-RFModel size=2.6M, Domain=Learnable latent domain-based2021.10 | 18.9 | — | |
| DPRNN-TasNetModel size=2.6M2019.10 | 18.8 | 19 | |
| DPRNN#params=3.6M, training=from original paper2020.02 | 18.8 | — | |
| DPRNN-TASNETModel size=2.6M, Domain=Learnable latent domain-based2021.10 | 18.8 | 19 | |
| DPRNNParams (M)=2.62023.09 | 18.8 | — | |
| FurcaNeXt#params=51.4M, training=from original paper2020.02 | 18.4 | — | |
| Deep CASAModel size=12.8M2019.10 | 17.7 | 18 | |
| Deep CASAModel size=12.8M, Domain=Frequency domain-based2021.10 | 17.7 | 18 | |
| SkiM-KS8Model size=5.9M, GMACs=4.92022.10 | 17.4 | 17.8 | |
| DTCN+DMModel size=3.6M, GMACs=3.72022.10 | 17.2 | 17.4 | |
| SuDORM-RF 1.0x++Model size=2.7M, GMACs=2.32022.10 | 17 | — | |
| SuDORMRFParams (M)=2.72023.09 | 17 | — | |
| E2EPFDomain=Learnable latent domain-based2021.10 | 16.9 | 17.3 | |
| DTCN+SW+DMModel size=1.3M, GMACs=3.72022.10 | 16.1 | 16.3 | |
| DTCNModel size=3.6M, GMACs=3.72022.10 | 15.6 | 15.9 | |
| Conv-TasNet (w/o SC)Model size=3.4M, GMACs=3.52022.10 | 15.4 | 15.7 | |
| SuDORM-RF 0.5xModel size=1.4M, GMACs=1.22022.10 | 15.4 | — | |
| Conv-TasNet-gLNModel size=5.1M2019.10 | 15.3 | 15.6 | |
| Sign Prediction NetModel size=55.2M2019.10 | 15.3 | 15.6 | |
| ConvTasNet#params=5.1M, training=from original paper2020.02 | 15.3 | — | |
| Conv-TASNETModel size=5.1M, Domain=Learnable latent domain-based2021.10 | 15.3 | 15.6 | |
| Conv-TasNetModel size=5.1M, GMACs=5.22022.10 | 15.3 | 15.6 | |
| ConvTasNetParams (M)=5.12023.09 | 15.3 | — | |
| Tiny-SepformerS-32Model size=5.3M2022.10 | 15.2 | 16 | |
| DTCN+SWModel size=1.3M, GMACs=3.72022.10 | 15 | 15.3 | |
| Bi-LSTM-TASNETModel size=23.6M, Domain=Learnable latent domain-based2021.10 | 13.2 | 13.6 | |
| IBMtraining=from original paper2020.02 | 13 | — | |
| IRMtraining=from original paper2020.02 | 12.7 | — | |
| WA-MISI-5Model size=32.9M2019.10 | 12.6 | 13.1 | |
| Chimera++Model size=32.9M, Domain=Frequency domain-based2021.10 | 11.5 | 12 | |
| TasNet#params=32.0M, training=from original paper2020.02 | 11.2 | — | |
| CBLDNN-GAT#params=39.5M, training=from original paper2020.02 | 11 | — | |
| DPCL++Model size=13.6M2019.10 | 10.8 | — | |
| DPCL++#params=13.6M, training=from original paper2020.02 | 10.8 | — | |
| DPCL++Model size=13.6M, Domain=Frequency domain-based2021.10 | 10.8 | — | |
| TDANetParams (M)=2.32023.09 | 10.8 | — | |
| ADANet#params=9.1M, training=from original paper2020.02 | 10.5 | — | |
| ADANetModel size=9.1M2019.10 | 10.4 | 10.8 | |
| FurcaNeXtModel size=51.4M2019.10 | — | 18.4 | |
| FurcaNeXtModel size=51.4M, Domain=Learnable latent domain-based2021.10 | — | 18.4 | |
| FurcaNeXtModel size=51.4M2022.10 | — | 18.4 | |
| UPIT-Bi-LSTM-STModel size=92.7M, Domain=Frequency domain-based2021.10 | — | 10 | |
| uPIT-BLSTM-STModel size=92.7M2019.10 | — | 10 |