Pierre Badin
Denis Beautemps
Rafael Laboissiere
Inst. de la Commun. Parlee, 46 Ave. Felix Viallet, 38031 Grenoble Cedex, France
An extension of Heinz and Stevens' (alpha)(beta) model [Proc. of 5th Int. Congr. Acoust. A44 (1965)], where (alpha) continuously depends on both midsagittal distance and VT position (along the midline), has been optimized from x-ray cineradiographs and formant frequencies obtained on a set of vowels and fricatives uttered by a subject. The purpose of the study was to recover the midsagittal function (a vector of 50 midsagittal distances) of fricative consonants from the measured formant frequencies, in [VCV] contexts. In order to reduce the excess degrees of freedom of this ill-posed problem, the proposed model has been used as a major way to constrain an optimization algorithm based on a gradient descent technique. The gradient of the error between actual and desired formant values is computed through a back-propagation network implementing both sagittal-to-area conversion and acoustic wave propagation. Preliminary results are presented and discussed. [Work partly supported by CNRS & CEC-ESPRIT project SPEECH MAPS.]