Application of an optimization technique to the inversion of an articulatory speech production model

A procedure is proposed to identify parameters specifying the vocal-tract area function. Conventionnal optimization techniques fail to perform such a task because the articulatory-to-acoustic transformation contains severe non-linearities. So the problem is inevitably complicated by a set of local minima of the matching criterion. This paper describes a technique combining a well-adapted least-square algorithm and a table-lookup procedure which takes into account the non-linearities of the model function. Preliminary results on synthetic vowels are presented.