Classification of dysarthric and non-impaired speech based on prosodic features

Abstract

Prosodic differences between dysarthric and healthy speakers were studied. Six acoustic properties that are plausibly more influenced by suprasegmental aspects of speech (e.g., emphasis) than the segmental details of the words were measured. The time course of these properties were analyzed over each utterance by fitting Legendre Polynomials. The resultant Legendre coefficients were then fed to linear- and quadratic-discriminant classifiers. All of the six properties were individually capable of distinguishing dysarthric speech from healthy speech. Based on one acoustic property measured over a single short sentence, we could correctly classify a speaker as healthy or dysarthric 55–75% of the time, depending on the acoustic property used. More complex classifiers that used all the acoustic properties correctly classified the speaker 97% of the time based on nine utterances. The strongest difference between normal and dysarthric speech was in loudness. Dysarthric speakers did not reliably produce the loudness patterns associated with stressed syllables. They also had a wider range in amplitude, greater voicing variability, smaller excursions of fundamental frequency, and less final lengthening compared to healthy speakers. The classification we demonstrated may be extended to a become a graduated measurement of severity, thereby contributing to diagnostics and intervention in dysarthria.