Suitability of speaker normalization procedures for classifying vowels produced by speakers with dysarthria

Abstract

Speaker normalization, a process whereby the perceptual system of a listener recalibrates to accommodate individual speakers, is proposed to account for the ease with which we understand speech produced by multiple speakers with different sized and shaped vocal tracts. A variety of vowel-, formant-, and speaker-intrinsic or extrinsic transforms have been proposed to model speaker normalization of vowels produced by multiple speakers (e.g., Mel, Bark, and Lobanov methods). Suitability of such normalization procedures has been examined extensively in non-disordered speaker populations. Unknown at this point, however, is the appropriateness of normalization procedures for transforming spectrally distorted vowels produced by speakers with dysarthria. Thus, we examined the suitability of two transforms, Bark and Lobanov, for normalizing vowels produced by a heterogeneous cohort of 45 speakers with dysarthria. Non-normalized (Hertz) and Bark transformed vowel tokens were classified via discriminant function analysis (DFA) with 55% and 56% accuracy, respectively. Classification accuracy of vowel tokens normalized using Lobanov’s method was 65%. The results of the DFAs were compared to perceptual data, which revealed listeners identified vowel tokens with 71% accuracy. These results suggest vowel-extrinsic and formant- and speaker-intrinsic normalization methods (e.g., Lobanov) are better suited to model speaker normalization of dysarthric vowels.