Speech analysis and modelling using a sequential ARMA estimation technique

Pole/zero modelling of speech has - been investigated for many years, yet the techniques developed so far have not proven satisfactory for most speech processing applications. Often these techniques separate the modelling problem in two: first and attempt is made to estimate the vocal tract impulse response; then an ARMA model is fit to such impulse response. The purpose of this paper is to discuss a different approach which is best-suited for speech analysis. This approach consist in first deriving from the speech signal a running approximation of a pseudo glottal excitation. Using a sequential minimization technique, the parameters of the ARMA model are then derived to fit a system whose input is the excitation signal and whose output is the speech signal.