Algorithm info: sampling_rate: 44100
num_channels: 2
n_fft: 4096
window_length: 4096
hop_length: 1024
stft_loss_n_fft: [256, 512, 1024, 2048, 4096, 8192, 16384]
num_dual_path_blocks: 12
num_time_transformer: 1
num_freq_transformer: 1
transformer_type: latent
chunk_length: 0
predict_method: mask # direct, mask
mix_ln: false
feature_conversion: true
hidden_size_list: [4, 32, 64, 128]
band_SR: [0.175, 0.392, 0.433]
band_strides: [1, 4, 16]
band_kernels: [3, 4, 16]
batch_size: 8
patch_len: 66150 # 44100 * 1.5
Metrics:
Metric sdr for piano: -0.9982
Metric si_sdr for piano: -4.4529
Metric l1_freq for piano: 34.2021
Metric log_wmse for piano: 10.0194
Metric aura_stft for piano: 4.2051
Metric aura_mrstft for piano: 4.8715
Metric bleedless for piano: 9.9902
Metric fullness for piano: 13.4602
Metric sdr for other: 9.4281
Metric si_sdr for other: 9.0673
Metric l1_freq for other: 21.0039
Metric log_wmse for other: 8.3727
Metric aura_stft for other: 6.1539
Metric aura_mrstft for other: 9.9090
Metric bleedless for other: 30.2889
Metric fullness for other: 16.5628
Date added: 2025-02-22 |
