Cepstral analysis has a pretty good statistical basis (no pun intended) away from the source/filter model. There's a bit of magic in the MFCC computation where

Cepstral analysis has a pretty good statistical basis (no pun intended) away from the source/filter model. There's a bit of magic in the MFCC computation where you apply the discrete cosine transform (DCT). That's all about reducing correlation between components in the cepstrum and makes no sense unless you "get" the way the change of basis has high energy compaction (more information in fewer values). However this has nothing to do with a physiological understanding of human speech.