VT AREA FUNCTION MODELS AND INVERSION

A parametric model of area functions for vowels and consonants has been improved to allow for a more realistic modeling of the vowel [ u] and to produce more realistic prototypes of [ ae] modeled altematively as a front or a back vowel. Inversions from formant frequencies to area function parameters are performed in two stages. The first is a codebook lookup and the second is an optimisation procedure. Present experience is limited to vocalic area functions. Our present system provides a rapid conversion of formant frequency data to VT parameters and has provided promising results for short sentences. Un modele parametrique de Ia fonction d'aire pour !es voyelles et !es consonnes a ete ameliore afin de permeure une modelisation plus realiste de Ia voyelle [ u} et de produire desprototypesplus realistes du [ae] modelise aussi bien comme une voyelle avant que comme une voyelle arriere. L 'inversion des formants vers !es parametres de Ia fonction d'aire est realisee en deux etapes. La premiere est une exploration de dictionnaire, et Ia deuxieme une procedure d'optimisation. Notre experzence presente est limitee aux fonctions d'aires vocaliques. Le systeme actuel permet une conversion rapide des frequences de formants en parametres du conduit vocal, et a fourni des resultats prometteurs pour des phrases courtes. lntroduction The revised vocal tract model The basic principle is to decompose any vocal tract area function into two parts, an overall vocalic part with parameters settings appropriate for the particular coarticulation and a parametrically specified consonantal part which substitutes or modifies parts of the vocalic area function. The modeling of the VT area functions employed the three traditional independent parameters, Xc and Ac for constriction location and area and a lip parameter lr/A0 [1][2]. The detailed area function was constructed by a concatenation of 6 successive segments and a piriformis cavity shunting the outlet of the Iarynx tube. The constants describing the shapes and sizes of these segment were given default values uniquely determined by the three main control parameters. Some of these constants can be released and given the status of shape parameters for adjustment of the model to a specific speaker. The consonantal part of a VT area function is specified by four parameters. Two of these, analogous to Xc and ~ of the vocalic model, have the primary function of specifying the location and degree of consonantal constrictions. The two additional parameters, pertaining to the effective length and shape of the consonantal constriction are dependent on the primary consonantal and vocalic parameters [3]. Temporal organisation will involve the control of covarying and in part dependent vocalic and consonantal area function parameters subject to articulatory constraints and representative time constants.