Sub Roformer (vocals, instrumental)
Sub Roformer architecture is based on paper SubRoformer: Subband Time-Domain Representation for Music Source Separation.
Quality table
Detailed statistics on the Multisong dataset:
| Model | Vocals fullness | Vocals bleedless | Vocals SDR | Vocals L1Freq | Instrum fullness | Instrum bleedless | Instrum SDR | Instrum L1Freq |
| Sub Roformer (SDR vocals: 12.23) | 19.46 | 35.98 | 12.23 | 41.46 | 29.99 | 47.61 | 18.54 | 41.90 |
| Sub Roformer high fullness by becruily | 25.37 | 30.54 | 12.13 | 41.93 | 37.65 | 40.72 | 18.45 | 42.13 |