A fast admissible method for identifying a short list of candidate words

Abstract

We present a fast method for identifying a short list of candidate words that match well with some acoustic input to serve as an initial matching stage in a large vocabulary isolated speech recognition system that uses hidden Markov models (HMM) and maximum likelihood decoding. This method is admissible in that the short list returned by it is guaranteed to contain the maximum likelihood word given by the HMMs. Given HMMs for all the words in the vocabulary we derive a class of algorithms that are faster than a detailed likelihood computation using these models by constructing an estimator of the likelihood. Using such an estimator we produce a list of candidate words that match well with the given acoustic input. We describe several possible estimators that can be used to construct such an admissible algorithm. © 1992.