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 a comparison on table 1 of this paper : mi.t.u-tokyo.ac.jp/assets/publication/… $\endgroup$
    – mprouveur
    Oct 1, 2020 at 10:33
  • $\begingroup$ My intuition is that it is easier to get good results with hand engineered features but it could be time consuming and you are limited by how much you can test yourself. However if you have enough training data you can use NN to learn these by themselves and it may find some better than the ones you would engineer. Testing is always the best way to get an answer though ^^ $\endgroup$
    – mprouveur
    Oct 1, 2020 at 10:35


Your Answer

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