A comparative study of variability in landmark sequences and implications for dysphonic speech analysis

Abstract

Dysphonia negatively affects speech intelligibility especially in the presence of background noise. This may be because dysphonic speech often contains a higher proportion of noise and/or a lower proportion of harmonic power, leading to reduced information in the speech signal. Landmark (LM) analysis was designed to identify patterns of information in the speech signal that are particularly salient for the auditory system. Consequently, it describes speech as a sequence of LMs. Past studies successfully differentiated disordered speech from normal speech based on the number of times each LM occurs. While the count was a sufficient measure for their purposes, transitional patterns in LM sequences may yield more descriptive information on underlying mechanism of the intelligibility deficits. Shannon’s Entropy and Markov chain model were used to evaluate the difference in LM sequences between normal and dysphonic speech. Landmarks were obtained from the first sentence of the Rainbow Passage for 33 normal speakers and 36 dysphonic speakers using SpeechMarkTM software package. The variability in the transitional patterns of LM was significantly less in dysphonic speech compared to normal speech. This suggests intelligibility deficits may be due to the greater acoustical constraints inherent to dysphonic speech.