The suitability of Hidden Markov Model (HMM) for representing speech signal is discussed. Initially, HMM is introduced as a second order statistics extension of static/composite sequential reference pattern (as in DTW). The need for modeling a single sound as a sequence of states is illustrated by highlighting phonetic context dependent variation of spectral properties evident in spectrogram.
The slides are at
http://iitg.ernet.in/samudravijaya/tutorialSlides/ASR_MFCC_DTW_HMM_GMM_LM.pdf
