shenbuguanni

What if I test these algorithms with real data?Real data can not directly obtain the noise signal, is there such an algorithm code?

nn split

1

the way (parameters) of this subband is trainable or untrainable?

Hi, thank you very much for your work, I recently found the following problems while generating rir using hybrid method: 1. The distance from the simulated sound source to the...

Postfiltering

1

I can not find where is the code for Postfiltering.

train stage

21

Can the training stage directly use this forward code to train the model? What is there to pay attention to? Thanks!

loss问题

2

1. calloss_magmse全带幅度谱loss最后是除以batch和频点维度，而分段幅度谱loss(calloss_magmse_subband)除的是batch和帧数，因为output_mag.shape[2]应该是T维度； 2. 另外请问为什么不在T维度求平均呢？一般使用F.l1_loss的reduction直接用Mean就会在batch和F和T求平均，这样是对效果有啥影响吗？

In the fullband corpus, a lot of audio was missing, for example, some had only mic without ref, or both existed but the mic signal was empty!

Why are all real datasets downloaded at 16k sampling rates. Where is the 48k training set