To compare the MFCC methods, we propose to generate
data from two known theoretical distributions, then build a classifier
to classify the data. Knowing the theoretical distributions of the data,
we can validate the projected PDFs against
the known PDFs and can even compare classification performance
with the optimal Neyman-Pearson classifier.
To make the problem interesting,
we base one theoretical model on MFCC and the other on an
circular auto-regressive process, so we can compare MFCC features with AR features.
To make the problem even more interesting, we make the two models
as close as possible to each other - approximating the same
Comparison of MFCC approaches