Evaluation of a Word Recognition System Using Syntax Analysis

A speech recognition system composed of a programmed syntax analyzer and a speaker-dependent word recognizer has been evaluated. The system accepts complete sentences in which the successive words are spoken distinctly and in isolation. The purpose of the experiment is to determine the capability of syntax analysis for improving the accuracy of word recognition and for expanding the command ensemble of a voice-actuated system. The word recognition system, designed by Itakura, 1 is based on representing speech utterances by equally spaced frames of LPC coefficients. Recognition ensues from a comparison of a sample input pattern of LPC coefficients with an ensemble of stored reference patterns previously established by the designated speaker. The comparison consists of a frame-by-frame scan of a sample pattern against each reference pattern. A distance metric (or measure of dissimilarity) is calculated and accu1619 mulated by a dynamic programming technique as the scan proceeds. The vocabulary item corresponding to the reference pattern with the lowest accumulated distance is designated the recognized item. In addition, a distance rejection threshold is imposed. If the accumulated distance exceeds the threshold at any frame during a reference scan, that particular reference comparison is aborted. If all reference comparisons for a sample pattern are aborted, the result is said to be "no match" or "reject."