Reconstruction of human voice using parallel structure transfer function and its estimation error

In this paper, we present a new method to reconstruct the speech signal using the parallel structure transfer function. In the case of a speech signal, the characteristics of the glottis wave and the resonance in vocal and nasal tracts are convoluted. Specifically, because the glottis waves are not the ones generated by the resonance, the waves have a non-linear characteristic. In our method, the resonance of vocal and nasal tracts is reconstructed using an ARMA model transfer function. Furthermore, the glottis wave and the estimation error of ARMA transfer function are recorded by an MA model transfer function. The MA model transfer function prevents the information reduction. In this research, the number of coefficients of the MA model transfer function is reduced and a method of achieving information compression is presented.