Algorithm info: Trained by WSF, I use Musdb18hq to train MelFormer for a test for this model, which only output the vocals.
The MelFormer from KJ's pretrained model:
SDR Instrumental: 17.285107
SDR Vocals: 10.977636
The SCNet send before:SDR Vocals: 8.261752
SDR Drums: 10.776174
SDR Bass: 10.974363
SDR Other: 5.350250
Metrics:
Metric sdr for vocals: 8.3253
Metric si_sdr for vocals: 7.3707
Metric l1_freq for vocals: 32.9754
Metric log_wmse for vocals: 11.6624
Metric aura_stft for vocals: 7.0406
Metric aura_mrstft for vocals: 7.7424
Metric bleedless for vocals: 29.0026
Metric fullness for vocals: 13.4600
Date added: 2024-11-27 |
