Surface Modeling of Vocal Tract Shapes in Transition Segments of Vowel-consonant-vowel Syllables for Estimation of Place of Closure

Production of vowel-stop consonant-vowel syllables involves movement of articulators from the articulatory position of the vowel towards that of the stop closure to that of the vowel. Movement of articulators before and after a stop closure is characterized by formant transitions. The closure portion of a stop have zero or low signal energy and relevant spectral information is not available. Hence, LPC based estimation of vocal tract shape fails during stop closure. A technique has been investigated for estimation of place of stop closure by surface modeling of estimated area values during vowel-consonant and consonant-vowel transition segments, on the assumption that articulatory movements have low order dynamics. Modeling of area values was based on least-squares conic and cubic bivariate polynomials and Delaunay triangulation based surfaces. The technique was applied for estimation of place of closure for bilabial, alveolar, and velar stops, both unvoiced and voiced, in syllables of the type /aCa/ spoken by five speakers. Results were more consistent with conic surface based interpolation, than that of cubic surface and Delaunay triangulation surface based interpolation. The proposed technique can be used for improving effectiveness of speech-training systems for production of stop consonants by providing visual feedback of place of closure. INTRODUCTION A good knowledge of the articulatory-acoustic relationship is useful for better understanding of articulation, speech synthesis and recognition, diagnosis of speech disorders, and speechtraining for the hearing impaired persons. Shape of the vocal tract can be specified by its crosssectional area as a function of position along the tract length [1]. Estimation of the vocal tract area function from speech signal, an inverse problem, can be carried out by one of the several techniques: LPC analysis [2], use of formants and factor analysis [3], use of formants and perturbation theory [4], mapping via articulatory codebook [1], etc. Most of these techniques are reported to work satisfactorily for vowels. However, shape estimation fails if spectral information is not available, for example during stop closure duration [1], [2], [5]. Speech-training systems providing visual feedback of articulatory efforts, not visible from outside, are reported to be useful for improving vowel articulation [6] [10]. Most of these systems are based on visual feedback of vocal tract shape estimated by LPC analysis of speech. However, similar systems for speech training are not available for consonant articulation. Hence, for improving the effectiveness of speech-training systems, it is important to investigate technique for vocal tract shape estimation for consonant articulation. This paper presents a technique for estimation of place of constriction during stop closures of vowel-stop consonant-vowel (VCV) syllables. Production of VCV syllables involves movement of articulators from the articulatory position of the vowel towards that of the stop closure to that of the vowel. Movement of articulators before and after a stop closure is characterized by transitions in vocal tract shapes as well as formants. The proposed technique is based on surface modeling of estimated area values during vowel-consonant (VC) and consonant-vowel (CV) transition segments, and its two-dimensional (2D) interpolation during stop closure for estimation of place of constriction.