|
Article Excerpt
The present study sought to characterize dysarthric speech in terms of acoustic landmarks. Landmark analysis provides a means to relate acoustic events to underlying articulatory behavior thereby allowing for comparisons between highly intelligible speech and dysarthric speech along a set of distinct acoustic parameters. Automatic landmark detection algorithms were utilized to extract acoustic landmarks from recordings produced by nine speakers with dysarthria and one control. Findings indicated that speakers with dysarthria not only produced expected acoustic targets at lower rates than the control, they also inserted unexpected landmarks at higher rates. Thus the dysarthric speech stream not only contains noisy acoustic information but also additional acoustic cues that may serve to confuse listeners. Additionally, these data highlight the notion that intelligibility is more than merely the result of accurate production of acoustic-phonetic targets. Rather, intelligibility scores resulted from the cumulative effects of precise, imprecise, absent, and superfluous articulation. The present study suggests the utility of automatic landmark analysis in developing personalized dysarthria treatment by specifying the the acoustic cues that a speaker produces accurately while also identifying cues that a speaker fails to produce or inserts unnecessarily. Implications of this work on designing semiautomatic diagnostic tools and computer-assisted interventions are discussed.
**********
Dysarthria is a motor speech disorder characterized by weak, slow, and/or uncoordinated movements of the musculature involved in speech production (Duffy, 2005; Yorkston, Beukelman, Strand, & Bell, 1999). Severely dysarthric speech commonly appears unintelligible to unfamiliar listeners; however, those familiar with the speaker are often able to comprehend with high accuracy (Deller, Hsu, & Ferrier, 1991). This observation implies that the speaker is producing acoustic cues that, while seemingly unintelligible to the unfamiliar listener, are capable of conveying information. Deller et al. (1991) hypothesized that the dysarthric speech stream not only contains noisy acoustic information but also additional acoustic cues that serve to confuse the listener. The current study sought to provide quantitative evidence for this hypothesis.
Stevens' Lexical Access from Features (LAFF) paradigm (Liu, 1995, 1996; Slifka, Stevens, Manuel, & Shattuck-Hufnagel, 2004; Stevens, 1992, 2002; Stevens, Manuel, Shattuck-Hufnagel, & Liu, 1992) was applied to perform automatic landmark detection. Based on distinctive feature theory, Stevens' model provides explicit definitions for landmarks, the acoustic correlates of articulator-free features (Chomsky & Halle, 1968; Jakobson, Fant, & Halle, 1952). Articulator-free features, which are also referred to as manner features, do not depend upon the position of the speech articulators. Instead, they provide a description of vocal tract constriction by classifying a speech segment as a vowel, glide, or consonant (sonorant or obstruent). Landmark detection provides a metric for comparing highly intelligible speech to dysarthric speech along a set of empirically derived acoustic features, thus serving as a lens for identifying accurate as well as inserted acoustic cues. Characterizing the differences between healthy and dysarthric speech in terms of vocal tract constriction also has clinical utility in that it relates acoustic events to underlying classes of articulatory behaviors.
METHODS
Nemours Database
Recordings used in this study consisted of productions from the Nemours Database of Dysarthric Speech (Menendez-Pidal, Polikoff, Peters, Leonzio, & Bunnell, 1996; Polikoff & Bunnell, 1999). This database contained recordings from 11 young males (in their twenties and thirties; exact ages were not documented) with dysarthria of varying severities secondary to either cerebral...
|
