Constructing Candidate Word Lists Using Acoustically Similar Word Groups

Abstract

In this correspondence we discuss a method for identifying a set of candidate words that match well with a given utterance. The method uses precomputed groups of acoustically similar words. Given a vocabulary we compute for each word a short list of words that are acoustically similar to it. These groups are constructed using training utterances recorded by several speakers. We then use these groups in an isolated utterance recognition system to construct candidate word lists during recognition. Our experiments show that the correct word appears on the short list of candidate words constructed during recognition using this scheme over 98% of the time. © 1992 IEEE