I'm trying to understand the methodology from section 4 in this report about audio classification but I have some trouble understanding how they pre-process the data (dsp is not my area).

  • $30$ seconds audio clips are available, sampled at $22050\textrm{ samples/s}$.
  • The STFT is computed "on $2048$ sample windows at a frame rate of $31.25\textrm{ frames/s}$"

Somehow this results in spectrograms of size $149\times 149$ which are then feeded to a neural network.

  • Does each spectrogram cover $2048$ samples of the audio clip?

  • If so, is then $149$ the number of segments that we divide each spectrogram in?

  • So that $2048/149 = 13.745$ is the number of samples that we use for computing the DFT in each segment? How can this be a non-integer?

    To my understanding the frame rate is about how much overlap there are between windows, in this case the overlap would be $22050/31.25 = 705.6 \textrm{ samples}$, is that correct? How can this be a non-integer?

  • $\begingroup$ "a report": which report? Cite! We can only guess from the numbers you're giving. Also, not all papers fully describe what they're doing, so this might simply be a case of "sorry, we don't know what the authors did, either". $\endgroup$ – Marcus Müller Apr 20 '17 at 9:55
  • $\begingroup$ Ok, added a link. But i think i gave all the relevant information. $\endgroup$ – termachine Apr 20 '17 at 10:12

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.