Speaker‐Independent Large Vocabulary Spoken Word Recognition by a Hierarchically Matching Method and a Method with Discriminant Functions

Abstract

AbstractThis paper reports on a spoken‐word recognition method for a large speaker‐dependent and speaker‐independent vocabulary employing a hierarchical matching technique. First, the two‐stage, speaker‐dependent method is discussed. In the first stage, the input pattern and the average standard are matched. Only in the cases where there are many similar words among the resulting high‐scoring candidate words is the second stage employed. The second stage employs a hierarchical matching method to match the high‐scoring candidates with a multipattern.When this method was applied to a large vocabulary containing 1000 words, the first‐stage recognition rate was improved 3 percent, thereby enabling a final recognition rate of 97 percent. Next, we report on the speaker‐independent method. Clustering is applied to standard word patterns derived from 40 speakers.A representative pattern from within the cluster and the multipattern which belongs to it are placed into a hierarchy. During the recognition phase, in the first stage, the input pattern and the representative pattern are matched.In the cases where there are many similar words among the high‐ranking candidates, the K” method is used in the second stage to match the multipattern which belongs to the representative pattern. Furthermore, in the final stage, a post‐processing method which empl.oys ten types of discriminating functions is used to redistinguish by automatically extracting the differing portions of similar words.Applying this hierarchical matching method to spoken‐word recognition for a 250/ 1011 word large speaker‐independent vocabulary, a high recognition rate of 94.5187.1 percent was possible.