SPEECH RECOGNIZING SYSTEM BASED UPON RELATIVE RELATION USING NEURAL NETWORK

摘要

PURPOSE:To enable highly accurate recognition processing by calculating the extent of matching between a couple of two kinds of vowels as the basis of recognition by the neural network as a relative position relation in the feature space between feature quantities obtained from the couple of vowels. CONSTITUTION:An optimum vowel candidate sequence recognizing means 104 finds the extent of matching between the couple of two kinds of vowel candidates as the output value of a neuron corresponding to the couple of the two kinds of vowel candidates of the output layer of the neural network 105 corresponding to the 1st example of the neural network 105. Namely, the neural network 105 calculates the extent of matching between the couple of the two kinds of vowels as the basis of the recognition as the relative position relation in the feature space between the feature quantities obtained from the vowels in the group. Therefore, the extent of the matching is calculated while hardly affected by the variation of a speaker and the influence of articulation coupling. Consequently, much more information regarding the relative relation between vowels than a system which represents the relative relation between vowels merely by distances between feature quantities is extracted and the highly accurate recognition processing is performed.