What is in the Dynamic Features: Analysis of the Derivatives of Log-Mel-Spectra

The present investigation analyses the behaviour of the first order derivatives of the log-mel-spectrum of vowels which constitutes the basis for the mel-frequency cepstral coefficients (MFCC). The results indicate that the dynamic features when inspected at log-mel-spectra level seem to be less influenced by speaker specific characteristics and degrade less in fast speech. However, when analysed in terms of MFCC the results indicate that the static features separate better between the vowel classes than the dynamic features. It could also be shown that the most important features for vowels with long duration occur at the beginning of a segment while for short vowels the most distinguishable point of measurement tends to be at the middle of the segment.