Analysis of voice source characteristics using a constrained polynomial model

This paper presents an analysis method of voice source characteristics from speech by simultaneously employing models of the vocal tract and voice source signal. The vocal tract is represented as a linear filter based on the conventional all-pole assumption. On the other hand, the voice source signal is represented by linearly overlapping multiple number of base signals obtained from a generalization of the Rosenberg model. The resulting voice source model is a polynomial function of time and has lesser degrees-of-freedom than the polynomial order. By virtue of the linearity of both models, the optimal values of their parameters can be jointly determined when the instants of the glottal opening and closing are given for each pitch period. We also present a temporal search method of these glottal events using the dynamic programming technique. Finally, experimental results are presented to reveal the applicability of the proposed method for several phonation conditions.