Speech analysis by pole-zero decomposition

Abstract

A new method of speech analysis based on a pole-zero decomposition technique is presented. The technique enables determination of pole parameters and zero parameters of a linear system model in an identical manner. Computation of the parameters is done in a straightforward recursive manner without the need for solution of either linear or nonlinear equations. The new method overcomes the limitations of the simplified all-pole models proposed for speech spectra. A pole-zero model for the vocal tract system not only provides an accurate description of the short-time spectral envelope but it also provides a means of obtaining accurately the source characteristics. The method has been applied to several difficult cases of real data such as voiced fricatives, nasals, transitions, etc., and the results indicate that the spectra derived from the pole-zero model match very well with the envelope of the short-time spectrum. It is shown that accurate and reliable voiced/unvoiced decision and pitch period estimation can be made by using the pole-zero decomposition technique.