about sdr about two_step_mask_learning HOT 2 OPEN

and the result calculated by pb_bss_eval

{'input_pesq': 1.5960057377815247,
 'input_sar': 11.243911897495014,
 'input_sdr': -0.5471245564855747,
 'input_si_sdr': -0.7163815595640699,
 'input_sir': 0.14099714373415928,
 'input_stoi': 0.662681954751808,
 'pesq': 2.596807837486267,
 'sar': 6.761317115018516,
 'sdr': 6.659954133819946,
 'si_sdr': 5.684012353271584,
 'sir': 23.99486035934128,
 'stoi': 0.8663816779000003}

from two_step_mask_learning.

Hey thanks for reaching out. I am kind of trying to catch a deadline so I did not check my github issues. It seems that your code is fine. Surprisingly there is a bug probably with the uploaded sources. If you listen to both the estimates and the actual sources you could actually hear that the sources and the mixture sound very noisy. However, the estimation seems to be quite better quality with much less artifacts. Moreover, I have actually used this code to produce some audio examples https://github.com/mpariente/asteroid/tree/master/egs/wham/TwoStep which might also contain this noise because of the wham dataset. I am gonna take a look at this, hopefully in a few days.

from two_step_mask_learning.