Quadratic classifier in nonstationary pattern recognition systems and its application to robust AR speech analysis

This paper is dedicated to an extensive comparative experimental analysis of nonstationary pattern recognition methods based on the quadratic classifier: the iterative quadratic classifications, its real-time modification, and the quadratic classifier with sliding training data set. The main purpose of this analysis lies in the choice of the nonstationary pattern recognition method based on quadratic classifier which gives the best results referred to the following criteria: a classification error, an adaptiveness in tracking of the nonstationary signals, and a sensitivity to the length of learning data set. The comparative analysis is done through analyzing the natural speech, isolated spoken Serbian vowels and digits, which are used as the examples of nonstationary signal