Determination of human vocal-tract dynamic geometry from formant trajectories using spatial and temporal Fourier analysis

This article presents a method of estimation of the vocal-tract cross-sectional area, considered as a function of time and position along the tract length. The estimation is based on the speech formant frequencies, and uses a priori information about natural tract configurations. In general lines, the method is as follows. First, the cross-sectional area is represented by a two-dimensional Fourier cosine series expansion in time and space. Then, the locally linear relationship between spatial Fourier coefficients and formant frequencies is explored to formulate an acoustical constraint in the coefficient space. Finally, the sequence of vocal-tract areas corresponding to a given sequence of formants is estimated under positional, dynamical, and acoustical constraints. The system behavior is shown first for the static case of vowels and, then, for the dynamic case of vowel-to-vowel transitions. The method can be used as a bridge between articulatory parameter models and the speech parameter space. Moreover, it is potentially useful for area driven coders and synthesizers