A methodology for efficiency estimation of the speech signal feature extraction methods

Abstract

This paper deals with a basis of methodology for speech signal feature description. Speech signal is described by three sets of features (the set of all descriptive features, the set of all selected features, and the set of all characteristic features). Feature description methods are described by three sets of maps (descriptive feature map, selected feature map, and characteristic feature map). As an example two feature description methods are considered — zero — crossing method and method of formant frequency energy classes (variant a and b). Efficiency of a single method being used in the recognition process has been estimated on the basis of experimental results. It is shown that the Fourier transformation as a map of descriptive features is more convenient as a measurement of time interval lenght. The mapping rule in variant b of the method of formant frequency energy classes gives a more convenient map of selected features than the mapping rule in variant a. With these maps the smallest features overlapping and consequently a better average recognition accuracy (greater than 92.5 %) can be achieved.