A method for segmenting acoustic patterns, with applications to automatic speech recognition

Abstract

Patterns derived by acoustic analysis of speech may be represented as a sequence of "segments", each described in terms such as length and mean spectrum. We present an automatic method of dividing the pattern from one utterance of a word into a sequence of segments. A "segment evaluation function" measures how well any given portion of the utterance can be represented as a single segment. The segmentation process itself simply finds the sequence of segments which best represents the whole pattern. Results are presented for several words and segment evaluation functions. Applications to automatic recognition of given words in connected speech are discussed.