Algorithm info: Single-model causal 4-stem Band-SCNet with 4 separation layers and 2.85M parameters. Trained on the combined MUSDB18-HQ and MoisesDB four-stem dataset. Best checkpoint at epoch 55. Strict causal inference with attention context 384 and block_frames=4. The instrumental estimate is the sum of drums, bass, and other. No ensemble or test-time model averaging.
Metrics:
Metric sdr for instrum: 13.9237
Metric si_sdr for instrum: 13.7163
Metric l1_freq for instrum: 29.3981
Metric log_wmse for instrum: 11.0304
Metric aura_stft for instrum: 9.8144
Metric aura_mrstft for instrum: 12.6299
Metric bleedless for instrum: 28.1942
Metric fullness for instrum: 24.8753
Metric sdr for vocals: 7.6856
Metric si_sdr for vocals: 6.7223
Metric l1_freq for vocals: 29.8112
Metric log_wmse for vocals: 11.0462
Metric aura_stft for vocals: 7.0596
Metric aura_mrstft for vocals: 8.5809
Metric bleedless for vocals: 16.8292
Metric fullness for vocals: 16.0456
Metric sdr for bass: 10.2703
Metric si_sdr for bass: 9.6973
Metric l1_freq for bass: 49.2957
Metric log_wmse for bass: 15.5491
Metric aura_stft for bass: 4.9921
Metric aura_mrstft for bass: 4.4895
Metric bleedless for bass: 20.8669
Metric fullness for bass: 17.2566
Metric sdr for drums: 9.7708
Metric si_sdr for drums: 8.6212
Metric l1_freq for drums: 31.5438
Metric log_wmse for drums: 13.8322
Metric aura_stft for drums: 6.6798
Metric aura_mrstft for drums: 6.3293
Metric bleedless for drums: 21.0536
Metric fullness for drums: 12.7580
Metric sdr for other: 5.0926
Metric si_sdr for other: 3.5686
Metric l1_freq for other: 28.8314
Metric log_wmse for other: 10.0861
Metric aura_stft for other: 5.6001
Metric aura_mrstft for other: 7.7217
Metric bleedless for other: 14.8732
Metric fullness for other: 16.4873
Date added: 2026-08-12 |
