search

fgnt / nara_wpe

Different implementations of "Weighted Prediction Error" for speech dereverberation
MIT License
474 stars 165 forks source link

speech enhancement with numpy WPE #63

Closed asadullah73-ce closed 2 years ago

asadullah73-ce commented 2 years ago

I can not see significant improvement in numpy WPE although I changed some parameters e.g delay -> 5 and iterations -> 15. But I think speech enhancement is not that significant. Please find the attached audio sample for improving. 20210909143636_26_395_sad_sam__5231.wav.zip

boeddeker commented 2 years ago

I am not sure, what exactly you mean with But I think speech enhancement is not that significant.. WPE is a blind dereverberation method. It only tries to suppress reverberation, but the effect on noise can be negative.

Regarding the parameters: WPE usually converges pretty fast. The difference between one and three iterations is often small. The delay has only a small effect. Not sure, but 5 may be to large. The number of channels and the number of taps are typically the parameters that you tune. More channels increase the effectiveness, while less channels can use/need more taps.

Your audio is a mono signal, so you will most likely only see a small effect, and you should increase the number of taps. Maybe 20, 30 or 40 will be better than the default.