Speech-training aids providing a visual feedback of articulatory efforts can be used for improving articulation by the hearing-impaired children. LPC-based estimation of vocal tract shape works satisfactorily for vowels but fails during stop closure. The vocal tract shape during the stop closures of vowel-consonant-vowel (VCV) utterances can be estimated by bivariate surface modeling of the vocal tract area function during the transition segments preceding and following the stop closure. The accuracy of the estimated shape during the closure depends on the detection of the transitions. A technique for detecting the VC and CV transitions in VCV utterances based on a measure of the rate of change of vocal tract area function is presented. The automatically marked start and end points of transitions showed a good match with the manually marked ones and resulted in a consistent estimation of the place of closure of velar, alveolar, and bilabial stops.

Index Terms: speech training aid, vocal tract shape, transition segment detection