Algorithm info: A custom model that adds register tokens to the attention layers, featuring Residual Flow Matching to decide to add-back residuals from BSRoFormer. FT from jarredou/BS-ROFO-SW-Fixed, with customized loss functions.
Metrics:
Metric sdr for instrum: 16.6535
Metric si_sdr for instrum: 16.5387
Metric l1_freq for instrum: 37.1092
Metric log_wmse for instrum: 13.4829
Metric aura_stft for instrum: 10.4792
Metric aura_mrstft for instrum: 12.9080
Metric bleedless for instrum: 44.7921
Metric fullness for instrum: 24.0210
Metric sdr for vocals: 10.3646
Metric si_sdr for vocals: 9.8554
Metric l1_freq for vocals: 36.9376
Metric log_wmse for vocals: 13.4830
Metric aura_stft for vocals: 8.5204
Metric aura_mrstft for vocals: 10.0972
Metric bleedless for vocals: 32.2166
Metric fullness for vocals: 15.6988
Metric sdr for bass: 13.2286
Metric si_sdr for bass: 13.0510
Metric l1_freq for bass: 56.8451
Metric log_wmse for bass: 17.8610
Metric aura_stft for bass: 5.9320
Metric aura_mrstft for bass: 5.2538
Metric bleedless for bass: 37.4566
Metric fullness for bass: 17.4862
Metric sdr for drums: 12.5482
Metric si_sdr for drums: 12.0461
Metric l1_freq for drums: 39.0524
Metric log_wmse for drums: 16.0058
Metric aura_stft for drums: 7.1315
Metric aura_mrstft for drums: 6.6450
Metric bleedless for drums: 38.1135
Metric fullness for drums: 12.9159
Metric sdr for other: 7.7370
Metric si_sdr for other: 6.8605
Metric l1_freq for other: 34.8281
Metric log_wmse for other: 12.4217
Metric aura_stft for other: 4.8728
Metric aura_mrstft for other: 6.8909
Metric bleedless for other: 25.3476
Metric fullness for other: 14.3743
Date added: 2026-05-02 |
