An approach to common acoustical pole and zero modeling of consecutive periods of voiced speech

Abstract

In this paper the open and closed phases within a speech period are separately modeled as acoustical pole-zero filters. We approach the estimation of the coefficients associated to the poles and zeros by minimizing a cost function based on the reconstruction error. The cost function leads to a matrix formulation of the error for two time intervals where the error must be defined. This defines a framework that facilitates to model the phases associated to consecutive periods. We give a matrix formulation of the estimation process that let us to attain two main objectives. Firstly, estimate the common-pole structure of several consecutive periods and their particular zero structure. And secondly, estimate their common-pole-zero structure. The experiments are carried out over a speech database of five men and five women. The experiments are done in terms of the reconstruction error and its dependence on the period length and the order of the analysis.