search

fgnt / nara_wpe

Different implementations of "Weighted Prediction Error" for speech dereverberation
MIT License
494 stars 164 forks source link

About WPE_Numpy_online: result question #77

Open gralear opened 6 months ago

gralear commented 6 months ago

Hi boeddeker I use the example WPE_Numpy_online and the wav from the test . but the dereverberated result have some problem. I try to change filter taps, but if the filter more but result get more bad.

Figure 2024-05-28 174318

boeddeker commented 6 months ago

Hi,

the idea of WPE is some kind of reducing the correlation between the current frame and the past. When you increase the number of taps, you increase the number of parameters. If you have too many parameters, it gets easier for WPE to predict the current frame. For offline WPE it can happen, that you get the zero signal, if the number of parameters matches the number of predicted time-frequency bins. The online variant has also issues, when you have too many parameters, but since I never use the online implementation, I cannot say, when it gets problematic.

gralear commented 6 months ago

Thank