Speaker-independent isolated word recognition using a 129-word airline vocabulary

Abstract

A reliable set of work reference templates for a speech recognition system can be obtained using sophisticated statistical clustering techniques. These studies have used a 39-word alpha-digit vocabulary and a 54-word vocabulary of computer terms. In this paper the clustering procedures, developed previously, are extended to create reference tokens for a 129-word vocabulary of airline reservation terms. Instead of using a single, casually obtained token of each word by each of 100 [human] talkers for the clustering procedure, each talker used the robust training procedure of Rabiner and Wilpon to provide a robust token of each vocabulary word. The effect on the clustering is to reduce the number of outlier tokens and to provide more robust clusters. Recognition tests using 20 new talkers yielded recognition accuracies of between 80%-97%. These accuracies are comparable to those obtained in an earlier study on the same vocabulary with speaker trained reference patterns.