Hello
I noticed that my corpus analysis (mfcc) is more efficient when setting minimum & maximum frequency to 600hz & 10000hz, for classification using kdtree. I’m about to record again the whole corpus, should I filter my audio signal while recording and apply the same filter when classifying or does it make no difference if I keep the whole spectrum?
I suppose it depends on the filters, but in general I’d avoid doing anything to the original audio (especially destructively) and let the internal filters deal with that, so that way you don’t have weird issues between center frequencies, different order of filters, phase funny business, diff q values etc…
In theory, you can filter or not, it makes no difference for MFCCs: it will only consider the spectrum between min and max in all cases. Now, in practice, no filter is perfect, so filtering and changing the range is likely to make a small difference.
I tend to record full range as I never know what I’ll try later, and it gives me more options.