Parameterized VT area function inversion

The purpose of our study is to contribute tools for inversion of articulatory to acoustics relations, in specific to perform an estimate of vocal tract area-function parameters from formant frequencies. The inversion is performed in two steps. A first approximation is attained from either a codebook or a neural net and a final optimization is performed by an iterative interpolation for finding a perfect or acceptable match. The study is based on a three-parameter vocal tract model. The codebook relates each of the possible combinations of constriction location, Xc, constriction area, Ac, and the lip parameter, l0/A0 to a corresponding F1, F2, F3 pattern. The neural network output provides the same choice of possible VT states as the codebook. The input to the neural network is normally programmed in terms of formant frequencies but other acoustic attributes can be selected or added. 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.