Adaptive learning algorithm in classification of fuzzy patterns: an application to vowels in CNC context

Abstract

An adaptive algorithm for recognition of ill-defined patterns using weak representative points and single pattern training procedure is presented from the standpoint of fuzzy set theory. The method includes both supervised and non-supervised schemes, A non-adaptive algorithm with fixed reference and weight vectors is also described to describe the efficiency of the system's adaptiveness to a new input. This was implemented to machine recognition of vowel sounds of a number of speakers in Consonant-Vowel Nucleus-Consonant (CNC) context considering the first three vowel formants as input features. The decision of the machine is governed by the maximum value of fuzzy membership function. A recognition rate, particularly for weak initial representative vectors, was seen to be dependent on the sequence of incoming patterns. As the process of classification continued, the learned moan vectors approached their respective true values of the clusters. Again, once the optimum size of training set is obtained, the role of the external supervisor became insignificant.