A polynomial approximation to the acoustic-to-articulatory mapping

An articulatory synthesiser has been developed that uses nine `quasi-articulatory´ parameters to represent the complete vocal tract outline. This outline is then sampled at regular intervals to obtain a set of vocal tract areas suitable for driving an articulatory synthesiser based on the Kelly-Lochbaum (1962) structure. Sets of parameters have been obtained for several vowels using magnetic resonance imaging. This data has been used to seed a collection of small codebooks, each designed to cover the region of acoustic space defined by a given voiced diphone. These codebooks have been filled using a parametric inching technique designed to obtain data points uniformly spaced in acoustic space. A least squares technique has been used to obtain a polynomial approximation for the acoustic-to-articulatory mapping. It has been found that a cubic mapping is sufficient to reproduce formant tracks to an acceptable degree of accuracy