Speaker-independent speech recognition using vocabulary-dependent clustering techniques

Abstract

The statistical technique of template clustering for speaker-independent speech recognition have typically used the same number of clusters to represent each word in the vocabulary. However, linguistic idiosyncracies and word confusability are contributing factors in determining the number of clusters required to adequately represent a given word. Based on this notion, an algorithm has been developed to assign the number of clusters for each word automatically, using a rule based on average cluster width. The objective is to reduce the computational load without undue degradation of recognition accuracy. Experiments have been carried out with a data base of 20 male and female speakers and a 40-word vocabulary that includes the alpha digits. The range of cluster levels was 3 to 8. Thus far, experimental results have shown that, with a constant cluster density and a total of 320 centroid templates, a recognition score of 82.75% was achieved with the alpha digits. At 291 templates allocated as a variable number of centroid templates per word, a recognition accuracy of 81 percent was achieved. Additional experiments with variations of the heuristic rules will be carried out to optimize performance.