0
$\begingroup$

I have a task to perform classification of audio signals using any suitable algorimth. After some research I found out, that CNN from this paper shows promising results. However, it still needs to be improved. Current data is scaled along each sample (that is each audio signals has zero mean and the same variance(i.e. 1)), and that leads to the loss of information about amplitude and variance of signal. Is there a way to append this information(i.e. if I compute it separately before scaling) to current signal to make CNN use it?

Thanks in advance for any help!

$\endgroup$
  • $\begingroup$ Can you quantify the loss of information? $\endgroup$ – grldsndrs Jul 31 '19 at 16:26
  • $\begingroup$ What kind of audio is this (music, env sounds, bird calls, ...)? What kind of classes are you interested in? $\endgroup$ – hendrik Aug 4 '19 at 8:01
0
$\begingroup$

As of 2019, the most common approach is to convert the raw audio waveform to a time-frequency representation ("spectogram"). The most commonly used spectrogram for general audio is log-scaled mel-spectrograms. This allows to use a 2D CNN, very similar to what one would use for images.

$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.