Sensitivity Based Segmentation and Identification in Automatic Speech Recognition

摘要

This research program continued an investigation of sensitivity analysis, and its use in the segmentation and identification of the phonetic units of speech, that was initiated during the 1982 Summer Faculty Research Program. The elements of the sensitivity matrix, which express the relative change in each pole of the speech model to a relative change in each coefficient of the characteristic equation, were evaluated for an expanded set of data which consisted of six vowels contained in single words spoken in a simple carrier phrase by five males with differing dialects. The objectives were to evaluate the sensitivity matrix, interpret its changes during the production of the vowels, and to evaluate inter-speaker variations. It was determined that the sensitivity analysis (1) serves to segment the vowel interval, (2) provides a measure of when a vowel is on target, and (3) should provide sufficient information to identify each particular vowel. Based on the results presented, sensitivity analysis should result in more accurate segmentation and identification of phonemes and should provide a practicable framework for incorporation of acoustic-phonetic variance as well as time and talker normalization. (Author)