Algorithm info: A custom model that adds register tokens to the attention layers, featuring a simplified Owned Calibrator inspired by exveria1015/Residual-Allocator (code is available on GitHub, without post-refining step). FT from jarredou/BS-ROFO-SW-Fixed, with customized loss functions.
Metrics:
Metric sdr for instrum: 16.4618
Metric si_sdr for instrum: 16.3422
Metric l1_freq for instrum: 36.7496
Metric log_wmse for instrum: 13.3125
Metric aura_stft for instrum: 10.6969
Metric aura_mrstft for instrum: 12.8165
Metric bleedless for instrum: 43.8348
Metric fullness for instrum: 23.8733
Metric sdr for vocals: 10.1705
Metric si_sdr for vocals: 9.6923
Metric l1_freq for vocals: 36.8854
Metric log_wmse for vocals: 13.3125
Metric aura_stft for vocals: 8.5316
Metric aura_mrstft for vocals: 10.2740
Metric bleedless for vocals: 30.1344
Metric fullness for vocals: 15.6418
Metric sdr for bass: 13.0651
Metric si_sdr for bass: 12.8865
Metric l1_freq for bass: 56.6384
Metric log_wmse for bass: 17.7092
Metric aura_stft for bass: 5.9122
Metric aura_mrstft for bass: 4.8543
Metric bleedless for bass: 36.9716
Metric fullness for bass: 17.4360
Metric sdr for drums: 12.4455
Metric si_sdr for drums: 11.9456
Metric l1_freq for drums: 38.8461
Metric log_wmse for drums: 15.9459
Metric aura_stft for drums: 6.4920
Metric aura_mrstft for drums: 5.7041
Metric bleedless for drums: 39.3368
Metric fullness for drums: 11.9554
Metric sdr for other: 7.4699
Metric si_sdr for other: 6.5307
Metric l1_freq for other: 34.3188
Metric log_wmse for other: 12.1742
Metric aura_stft for other: 4.9567
Metric aura_mrstft for other: 6.8613
Metric bleedless for other: 24.9781
Metric fullness for other: 13.7675
Date added: 2026-04-29 |
