Denoise Pretrained Model and norm_noisy.mat

[demeng]

Two questions here for some help:

1. We tried to use the pre-trained model /backup/TSN.pd to denoise the noisy audios. The results so far have large amount of added white noise in it. Here is an example: the input audio is
   noisy01_snr03_num0001.raw
   And the below enhanced audio has a lot of added white noise
   enhanced01_snr03_num0002.wav
   Have anyone used the pre-trained models to enhance noisy audios successfully? If so, can you point us what we might be doing wrong?

2. What norm_noisy.mat should be used for testing/inference? The code suggests the norm_noisy.mat is from the training data. If so, can you share the norm_noisy.mat for the pre-trained model?

Any advice is appreciated! @jtkim-kaist @MayMiao0923 @yangz-zju @ucasiggcas

[shamoons]

Same question

[jtkim-kaist]

1. Did you try to use the test code in test_small.py?

2. norm_noisy.mat is very important in this code, which means the mean & variance of whole training set. norm_noisy is used to reconstruct the wav from predicted spectrogram. You can make norm_noisy.mat from get_norm.py as described in readme. But if you cannot make cause of any reason, I will upload my own but it will be slightly different from the one of your training dataset.