<<Prev Up Next>>

What happened?

Let's take a look at sample 6, the sample that faired the worst.  You can listen to the sample here.

Spectrogram of test sample 6

This spectrogram looks quite different from the previous one.  Music to the left of the black line is solo and to the right is non-solo.   Let's take a look at what the classifier did.

Classification of test sample 6

Well, it seems as though the classifier had two problems: (1) it mis-classified the beginning as non-solo and (2) it mis-classified a lot of the non-solo section as solo.  Let's tackle (1) first.  At the beginning of the solo section, the pitch is quite low and so the harmonics are quite close together.  That leaves less space for zeros in between the harmonics and so these time slices did not meet the cutoffs.

In the non-solo section, just take a look at the spectrogram again.  There's a lot of space between harmonics in the 1000-400 Hz range.  If you listen carefully to the sample, you will hear the orchestra come in rather quietly under the soloist.  I think that is what is being reflected in the spectrogram.  Unfortunately,  this obviously threw off the zero count.

<<Prev Up Next>>

Christine Smit

Christine Smit's email address