Algorithm info: FourSepNet SqrtMC v1.0.
Single 210K param model, not an ensemble.
Checkpoint: best epoch 107.
Input: 44.1 kHz stereo audio.
Output stems: vocals, drums, bass, other and instrumental.
Instrumental is calculated as drums + bass + other.
STFT: n_fft=4096, hop_length=1024.
Inference framework: PyTorch 2.0.1, CUDA 11.8.
Inference device: NVIDIA GeForce RTX 4090 D.
Metrics:
Metric sdr for instrum: 12.9526
Metric si_sdr for instrum: 12.6979
Metric l1_freq for instrum: 28.3433
Metric log_wmse for instrum: 10.1279
Metric aura_stft for instrum: 11.3814
Metric aura_mrstft for instrum: 11.6418
Metric bleedless for instrum: 26.4346
Metric fullness for instrum: 24.6108
Metric sdr for vocals: 6.6452
Metric si_sdr for vocals: 5.5248
Metric l1_freq for vocals: 27.8884
Metric log_wmse for vocals: 10.1279
Metric aura_stft for vocals: 6.6303
Metric aura_mrstft for vocals: 7.9544
Metric bleedless for vocals: 14.8270
Metric fullness for vocals: 15.4222
Metric sdr for bass: 9.4361
Metric si_sdr for bass: 8.8676
Metric l1_freq for bass: 47.1358
Metric log_wmse for bass: 14.8900
Metric aura_stft for bass: 4.8919
Metric aura_mrstft for bass: 4.3495
Metric bleedless for bass: 21.7411
Metric fullness for bass: 17.9116
Metric sdr for drums: 8.8748
Metric si_sdr for drums: 7.8432
Metric l1_freq for drums: 30.7533
Metric log_wmse for drums: 13.2015
Metric aura_stft for drums: 6.8774
Metric aura_mrstft for drums: 6.0599
Metric bleedless for drums: 17.4490
Metric fullness for drums: 13.9924
Metric sdr for other: 4.2056
Metric si_sdr for other: 1.8920
Metric l1_freq for other: 26.0266
Metric log_wmse for other: 9.2126
Metric aura_stft for other: 4.5246
Metric aura_mrstft for other: 6.2139
Metric bleedless for other: 14.3052
Metric fullness for other: 13.5293
Date added: 2026-08-28 |
