Speaker-independent word recognition in connected speech on the basis of phoneme recognition

Abstract

A method of speaker-independent connected-word recognition by robust segmentation for speaker variation is described. To normalize the variation by speakers, an input speech pattern is transformed through segmentation and labeling into a sequence of phonemically labeled segments (phoneme string) which have less variation by speakers. Connected word recognition is carried out using a two-level DP matching algorithm on that phoneme string. The input speech pattern is oversegmented in order to avoid omissions which cause fatal errors in word recognition. The number of segments which correspond to one phoneme should depend on the phoneme; the number of segments for vowels should be greater than that for consonants. From this viewpoint, we propose a method of varying the matching path adaptively with respect to each phoneme, at the dynamic-programming word-matching level. In experiments on spokenword recognition of one to four connected digits, the recognition rate for each word was about 90% and for each sequence of words was about 80%, on an average over seven male speakers. In the case where the words are spoken clearly, the former improved to 93.8% and the latter to 86.0% on an average.