A dynamic processing approach to extraction and categorization of phonemic information

This paper aims to report a new successive procedure for obtaining a sequence of phonemic categories from acoustic speech signal. The procedure is composed of three levels, i.e., enhancement and inhibition of speech spectra (preprocessing), local peak extraction (feature extraction), and processing of dynamic properties of the feature series (categorization). Each level is composed of multistage processes whose one stage consists of two or three computational steps. The principal features of this procedure are: (1) Formulations of contrast effect, assimilation effect and uncertainty relation between time and space. (2) Potential function defined for phonemic categories. (3) Lattice-like configuration of functional devices. (4) No feed back loop process and no back tracking algorithm. The feasibility of the procedure is verified by applying it to continuous vowels and words. The advantages are discussed from the viewpoints of reliability and flexibility of pattern recognition techniques.