I am trying to distinguish between speech signals and music signals using feature extraction and SVM. I am looking for features to include for this purpose.
While going through literature, I came across something called '4 Hz modulation energy' that is distinctively present in speech and not in music, but was unable to find any further literature about it. I would like to know the reason for this energy present in speech. References to the same would be very helpful.
Also, I tried plotting the spectrogram of a music signal and a speech signal, such that the frequency bin size was 1Hz (which would involve me taking a 44100 point FFT where 44100 was my sampling frequency). On visually inspecting the spectrogram at 4Hz, such a 4Hz peak of energy was not apparent to me for a speech signal as compared to a music signal. When I include the 4 Hz modulation energy as a feature per frame in my SVM, it indeed reduces the accuracy of prediction.
As such I would like to know how this energy arises, and would like to know more about it.
Edit : This paper answers some of my questions.