I want to compute kaldi pitch features at the end of a network. Is there a way to compute gradients of the kaldi pitch feature extraction? (Maybe something similar to spectrogram of torchaudio).
Thanks!
Hi,
I’m not familiar with kaldi pitch, how do you compute it? If you use pytorch construct, the autograd should work just fine 
The original Kaldi library is written in C and usually called using .sh files. I do not know of a Pytorch function that implements it, although there are python envelopes such as pykaldi.
If there is no pytorch implementation, I’m afraid you’ll have to do one of the following:
- reimplement a new version with pytorch operators to use the autograd
- write a custom Function (https://pytorch.org/docs/stable/notes/extending.html) that uses the original library in the forward and for which you will specify the backward.