Universal Sound Separation on MC-FUSS
18.9SI-SDRi (J=2)DeepASA (SEP)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DeepASA (SEP)Training=Fine-tuning (pre-trained on ASA2), Param.=2.9 M, MAC/s=77.3 G2025.09 | 18.9 | 19.1 | 17.6 | 18.5 | |
| DeepASA (SEP, w/o ATST)Training=Fine-tuning (pre-trained on ASA2), Param.=2.9 M, MAC/s=77.3 G2025.09 | 18.8 | 19 | 17.4 | 18.4 | |
| DeepASA (SEP, w/o ATST)Training=From scratch, Param.=2.9 M, MAC/s=77.3 G2025.09 | 18.5 | 18.2 | 15.7 | 17.5 | |
| DeFT-MambaTraining=From scratch, Param.=3.6 M, MAC/s=83.8 G2025.09 | 18.4 | 17.1 | 13.8 | 16.4 | |
| DeFT-MambaTraining=Fine-tuning (pre-trained on ASA2), Param.=3.6 M, MAC/s=83.8 G2025.09 | 18.4 | 17.1 | 14.1 | 16.6 | |
| SpatialNetTraining=From scratch, Param.=7.3 M, MAC/s=71.8 G2025.09 | 17.8 | 16.5 | 13.1 | 15.8 | |
| DeFTAN-IITraining=From scratch, Param.=4.1 M, MAC/s=66.1 G2025.09 | 17.6 | 16.3 | 12.8 | 15.6 | |
| TF-GridNetTraining=From scratch, Param.=14.7 M, MAC/s=462 G2025.09 | 17.2 | 16.1 | 12.5 | 15.3 | |
| MC-BSRNNTraining=From scratch, Param.=12.2 M, MAC/s=15.3 G2025.09 | 15.7 | 15.2 | 11.4 | 14.1 | |
| ByteDance-ussTraining=From scratch, Param.=28.0 (+80.7) M, MAC/s=40.1 G2025.09 | 14.8 | 14.4 | 12.7 | 14 |