Classification of accurate and misarticulated rhotic syllables for simplified ultrasound biofeedback therapy

Abstract

Ultrasound biofeedback therapy (UBT) provides real-time imaging of tongue movements and has demonstrated positive speech remediation outcomes; however, some individuals have limited or no response. UBT outcomes could be further improved by a simplified biofeedback display to enhance motor learning. Such simplification requires automatic processing of ultrasound images to determine biofeedback parameters and targets. We investigate potential biofeedback parameters using TonguePART, a method that automatically tracks the tongue surface on midsagittal ultrasound images to quantify displacement trajectories of the tongue root, dorsum, and blade. Our focus is rhotic syllables (/i/, /u/, /o/, /e/, /e/, and /a/ with initial or final /r/) from children with residual speech sound disorders and children with typically developing speech. We train support vector machines on measured tongue part displacement trajectories to distinguish between accurate and misarticulated productions as determined from auditory perceptual ratings. Preliminary data indicate that a linear combination of the tongue dorsum and blade displacements, between the vowel and consonant, can distinguish between accurate and misarticulated productions of rhotic syllables. These results suggest a real-time biofeedback parameter based on projections of real-time dorsum and blade displacements, along with potential target values, different for each vowel, for this parameter in simplified UBT for speech remediation.