I am working on classifying different sounds ( not speech or words exactly something like ambulance alarm, police alarm, cough sounds etc)

I read few paper which suggested to extract dsp features such as MFccs, skewness, kurtosis, log energy, entropy, zcr etc ( using total around 20 features ) from each segmented sound event.

I am currently using all these features and training xgboost, 3 layer DNN with ReLU, I am getting good accuracy.

I also read recent deep learning papers where they just used spectrogram images and feed it to convolution network which is capturing both temporal and spatial features. ( Using spectrogram one feature only )

I am looking for some explanation which one is better method for classifying sounds and why?

Any reference paper for comparison would be additional help.

Thank you!

  • $\begingroup$ You can find SOTA results in DCASE challenge. Each submission is graded based on F-score performance (and on this year also PSDS for polyphonic systems). You've got tables with features used, system architecture, etc. Everything that you asked for. $\endgroup$
    – jojeck
    Oct 1, 2020 at 8:10
  • $\begingroup$ @jojek can check this question? dsp.stackexchange.com/questions/70631/… $\endgroup$ Oct 1, 2020 at 8:13
  • $\begingroup$ Well, it's an old argument of hand-crafted features vs learned features. All the features that you've mentioned (MFCC's, skewness, kurtosis, etc.) are assumed to carry relevant information which isn't necessarily the case. Instead, you extract spectro-temporal map and allow the features to be learned by CNN filters. For most of the applications CNN will work better because they extract things that are optimal and can't possibly think of. As for your ZCR example, yes, spectrogram doesn't contain it but it's highly correlated with pitch and noise. $\endgroup$
    – jojeck
    Oct 1, 2020 at 8:21


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.