Application of a Landmark-Based Method for Acoustic Analysis of Dysphonic Speech

Abstract

Speakers with dysphonia often report difficulty with maintaining intelligibility in noisy environments; however, there is no objective method for characterizing this difficulty. Landmark-based analysis is a linguistically-motived, knowledge-based speech analysis technique, which may serve as the basis of acoustic tool for describing the intelligibility deficit. As the first step toward development of such a tool, this study examined whether Landmark-based analysis could describe acoustic differences between normal and dysphonic speech. The recordings subjected to the Landmark-based analysis were the first sentence of the Rainbow Passage from 33 speakers with normal voice and 36 speakers with dysphonia. These recordings were selected from the Kay Elemetrics Database of Disordered Voice. The between-group difference was evaluated based on counts of certain Landmarks (LM). The average counts of all LMs were significantly greater in normal speech, t(66.85) = 2.36, P = 0.02. When the group-difference was examined for each LM, dysphonic speech had more [g] and [b] LMs and fewer [s] LMs than normal speech (P < 0.01 for all cases). A classification tree model identified [+s] and [+b] LMs are the primary predictors for the dysphonic speech. The model's misclassification rate was 7.24%. This preliminary investigation demonstrates that LM-based analysis is capable of differentiating dysphonic speech from normal speech. This encouraging result rationalizes future examinations of LM analysis in other areas of interest. For example, LM-based measures could conceivably be used as to quantify general intelligibility, and/or provide insight into underlying mechanisms of intelligibility deficits.