I have audio data acquired from a 4 channels sensors array.

As a preprocessing step for a neural network, I want to beamform and focus on the sound source. For higher resolution in the beamforming process, I have upsampled the signal with a factor of 5 using cubic interpolation, from 48 kHz to 240 kHz.

I am aware that now my network will train much slower due to the longer vectors. Nevertheless, my network is not performing well (it did without preprocessing step). Is there any connection? Is there any reason to downsample the frequency back for machine learning purposes specifically and for classical signal processing method in general?

  • 1
    $\begingroup$ You should probably use FFT interpolation (FFT, zero-pad to 5 times the original length, then inverse FFT) rather than polynomial interpolate. This will provide more accurate interpolation for a band-limited signal. However, (I'm not an expert on NN), this may give you a clue why interpolation isn't helping - if the first layer of your NN is effectively FFT kernels, you have just introduced zeros (with FFT interpolation) or junk (with with polynomial interpolation). $\endgroup$ – mikado May 19 '19 at 19:40
  • $\begingroup$ @mikado is this achievable by simply using resample() method in python as in this link? $\endgroup$ – havakok May 20 '19 at 6:28
  • $\begingroup$ That is an implementation of the technique I suggest $\endgroup$ – mikado May 20 '19 at 18:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.