You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi, I have read LEAF and its previous version: time domain filterbanks, and I notice that if I want to approximate mel filterbanks with the proposed audio frontend, the support of the window function should be larger than the support of the gabor filters.
However it seems that it doesn't hold for all window_size, mel_bins setting. It the paper, 40/64 mel_bins are used, Currently, I want to use 80 mel_bins because this is usually used in self-supervised learning paper.
I'm wondering will it cause severe approximation error due to increased support of gabor filters? Is there any design principle for the determination of window_size, mel_bins combination to make sure no severe approximation error occur?
The text was updated successfully, but these errors were encountered:
Hi, I have read LEAF and its previous version: time domain filterbanks, and I notice that if I want to approximate mel filterbanks with the proposed audio frontend, the support of the window function should be larger than the support of the gabor filters.
However it seems that it doesn't hold for all window_size, mel_bins setting. It the paper, 40/64 mel_bins are used, Currently, I want to use 80 mel_bins because this is usually used in self-supervised learning paper.
I'm wondering will it cause severe approximation error due to increased support of gabor filters? Is there any design principle for the determination of window_size, mel_bins combination to make sure no severe approximation error occur?
The text was updated successfully, but these errors were encountered: