Incremental learning mechanisms for speech understanding

The feasibility of using machine-learning techniques to support the development of a speech understanding system is investigated. The research studies the behavior of hidden Markov models as an incremental learning mechanism. The approach is to train whole-word recognizers using a minimal set of training tokens and then incrementally train them as new examples of the words are recognized and added to the training set. The results show that it is practical to generate seed recognizers from a small training set made up of word tokens composed from segmental fragments. This yields better results than a model formed by chaining hidden Markov models. A seed recognizer can be used to locate additional tokens. One iteration of the Baum-Welch training algorithm is sufficient to incrementally train the hidden Markov model. There is a rapid improvement in the performance of the hidden Markov model with each new token added