MelBands malfunctions when number of bands is large (e.g. 128)

[justinsalamon]

MelBands (when fed with the output of Spectrum) seems to work fine for low numbers of bands, e.g. 24 or 40. But if we increase this number to e.g. 128 or 256, it seems to (mistakenly) always return 0 for specific band numbers.

For instance when setting numberBands = 256, regardless of which audio file I analyze I get 0 in the following bands:

array([[ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15, 16,
        17, 18, 19, 20, 21, 22, 24, 25, 26, 27, 28, 29, 30, 31, 33, 34, 36,
        37, 39, 40, 41, 42, 44, 45, 46, 47, 49, 51, 53, 54, 56, 58, 60, 62,
        64, 67, 69, 71, 74, 79, 82]])

Here are some example spectrograms (of the log-melbands, ignore the y-axis labels):

[justinsalamon]

update: I think the issue lies with the computation of the start/end spectral bins for each band - if the FFT size is relatively small (e.g. 1024) and the number of mel bands is large (e.g. 128) then for some bands the start and end bin will be the same bin (or adjacent bins), which consequently results in 0 energy being summed at that band.

A practical solution is to use zero padding to increase the number of FFT bins thus avoiding the problem. Nonetheless, it would probably be helpful to run a check when the mel filterbank is created to ensure there are no bands with adjacent or identical start/end bins, and if any are detected print out a warning.