Isolated word recognition using hidden Markov models

In this paper we investigated smoothing techniques for alleviating the problem due to insufficient amount of training data. Hidden Markov Models (HMM) require a large amount of training data to obtain reliable probability estimates. But for isolated word recognition (100 words or more), we can not expect a user to speak each word more than several times. We found that the confusion matrix between a pair of label prototypes was particularly effective for the problem. We investigated two ways of computing the confusion matrix. One is based on distance among labels, and the other is based on the correspondence of labels in several utterances of the same word. Performance of these techniques was tested by using 100 Japanese city names spoken in an isolated word mode by three speakers. It was found that the smoothing technique reduced recognition errors from 1% to 0.1%. To visualize such performance improvement, we used, together with recognition rate, "two-dimensional score plot," which shows the distribution of the best score of the true word and that of the remaining false ones in the vocabulary.