`numberOfMFCCCoefficients` vs `melBands` about meyda

[hughrawlinson]

I confused myself, was mixing up the barkBands field that stores the bark bands with the number of bark bands that should be computed.

[hughrawlinson]

I'm still confused lol

There's numberOfMFCCCoefficents, which is the number of mfcc coefficients that should be calculated.
There's melBands, which is the number of mel bands calculated. It's unclear to me if it's useful for this to be separate from numberOfMFCCCoefficients.
There's barkScale, which is unrelated to either of these two, and is related to loudness.
There is no barkBands.
#971 adds numberOfBarkBands to let the user specify how many bark bands to use.

So the question stands: why would we set numberOfMFCCCoefficients and melBands separately?

I'm mainly asking this so I can document it, I'm sure there's a reason.

[jamesb93]

I'm no expert but if you need 40 MFCC coefficients you need to start with at least 40 bark or melbands. MFCC is effectively a form of dimension reduction and the amount of information you start with before the liftering etc. will effect the result.

[hughrawlinson]

Yes, that’s true - but I think right now all we use it for is the mfcc feature so I’m not sure why we would ever need to calculate more mel bands than we would then need to use in that mfcc calculation.

[jamesb93]

Because it changes the MFCC result.

For example, doing some rudimentary tests in Max:

For a given a spectral frame the values are different, albeit slightly. I think it would be sensible to set a high default (40 bands) and to then have it adjustable. I know in my use case where I use Meyda for pedagogical purposes not having the technology assume what I want is useful and sensible defaults make it approachable for those who may be still learning.

[jamesb93]

Because it changes the MFCC result.

For example, doing some rudimentary tests in Max:

For a given a spectral frame the values are different, albeit slightly. I think it would be sensible to set a high default (40 bands) and to then have it adjustable. I know in my use case where I use Meyda for pedagogical purposes not having the technology assume what I want is useful and sensible defaults make it approachable for those who may be still learning.

To clarify: the MFCC on the left is using 40 melbands and the one on the right is doing 20 melbands before producing 13 MFCC coefficients in either case.

[hughrawlinson]

Aha! Thanks, that fully clarified it :)