Fast Match Based on Decision Tree

In a large vocabulary speech recognition system using hidden Markov models, calculating the likelihood of an acoustic signal segment for all words in the vocabulary involves a large amount of computation. We describe in this paper a scheme to rapidly obtaining an approximate acoustic match for all words in the vocabulary in such a way as to ensure that the correct word is one of a small number of words examined in detail. Using a decision tree method we obtain a matching algorithm that is much faster than common acoustic likelihood computation on all the words. This method has been tested on isolated syllables.