KinWaiCheuk

The mel filterbanks should update according to the gradient. Can you compare the gradient for the mel filterbanks between the first few epochs and the epochs when it stops updating?...

Hi @arachid1. Thanks for using nnAudio! The STFT kernel updates are gradient driven (and gradient are calculated from the loss function that you used). Just like other neural network layers...

Thanks for your update! Regarding the reason why the higher bins backpropagate more than the lower bins despite richer information in lower bin, I also have no idea. But it...

May I know which library are you using for `time_stretch` and `pitch_shift` now? As far as I know, these operations will output a modified version of the audio arrays. If...

Thanks for the suggestion. I have not used torchlibrosa before, but it seems it only supports STFT at the moment? I would like to know more about this to make...

Yes please! It would be a great help. Then we would also need to add a few more test cases into the [pytest file](https://github.com/KinWaiCheuk/nnAudio/blob/master/Installation/tests/test_stft.py#L79), if possible, to make sure the...

I think approximation errors should be good enough!

The learnable kernels are, by default, disable in all STFT, Mel spectrograms, MFCC, and their inverse counterparts. If you are referring to the argument `refresh_win` as shown below, it is...

> Thank you for the answer. I want to use RNNS instead of CNNs in my maodel. Will this code work for RNNs too? Yes it will work. nnAudio is...

May I know is there anything special about that batch that causes the slowness? I am unable to reproduce to same behaviour. What is the GPU you are using? How...