A Study of Vocal Tract Shapes Using Magnetic Resonance Imaging and Acoustic Reflectance Techniques

A major di culty in performing speech synthesis using an articulatory model is the lack of su icit data on the range of vocal tract shapes. The vocal tract is usually represented as a concatenation of short uniform tubes, the areas of which are referred to as the vocal tract area function. In order to produce synthetic speech, a sequence of these area mctions is required, which when used in the synthesiser, willmimic a give: natural utterance. A commonly used technique for obtaining a sequence of area rnctions employs a very large'code book, randomly populated with vocal tract shapes The task of searching such a code book is computationally expensive, though this cost can be somewhat reduced using a neural net mapping method 1. For copy synthesis of a single speaker however. a much more e icient code book could be generated if area rnctions close ₜo those used by the sneaker were known. In some earlier work 2, magnetic resonance imaging was used to determine the vocal tract area rnctions of a male speaker producing 5 steady state vowel sounds The data was used in an acoustic tube sinmlation, and the resdting acoustic spectra were compared with those of natural speech from the same speaker. For all but one of the vowels good agreement, typically within lZOHz, was obtained for the rst three formant equencies. Magnetic resonance imaging is however costly and time consuming. Moreover, a far larger number of shapes than could be obtained 'om imaging is needed for the task of acoustic to geometric mapping. For this reason, we investigated an acoustic technique, rst descn'bed by Soadhi and Gopinath 3. Preliminary results 4 showed that the basic acoustic technique, when used in conjunction with nther optimisation, provided a low cost method of extracting both rubber-model and real vocal tract area functions Further experiments have allowed us to study the changing vocal tract shape during diphthongs A parametric model of the vocal tract was employed to simulate the mechanism of the vocal tract during these transitions '