Absolute value of the output of sinc-based convolutions?

[mravanelli]

Absolute value of the output of sinc-based convolutions?

[hbredin]

Thanks for the clarification -- it would make a nice addition to the paper and definitely help others reproduce your experiment.

Here is my (slightly more pytorchic) version of SincConv that can be used as a drop-in replacement for torch.nn.Conv1d: https://gist.github.com/hbredin/c3bfa80f8f71181c2d499167d297f181

[mravanelli]

Hi Hervé,
we tried many different alternatives at this level and experiments are still running to find better solutions. For instance, we recently compared the performance with and without the abs function you pointed out. The reason why it could make sense to add an abs function here, is that standard max-pooling only looks into the positive samples of the filtered signal, while with abs we basically consider the envelope of it. This would allow the network to extract features more related to the “energy” or “envelope” of the signal, resembling “more closely” the standard pipeline used for extracting standard fbank or mfcc features. Actually, the standard fbanks is still quite different since we average the spectrum within rectangular filters. From my experiments, however, I saw that doing average pooling at this level works slightly worse than max pooling.

[mravanelli]

Great job! thank you!

[Rachine]

It seems that the addition of this abs operation is very important for fast convergence!