Recognition of word-final unstressed syllables

Abstract

A system was developed to recognize American English word-final unstressed syllables ending in a sonorant. The algorithm takes as inputs a digitized speech signal and two time pointers, one during the penultimate vowel and the other at the end of the input word. Output includes identification of the final segment and a list of features for the segment preceding the final segment. Identifying these segments is necessary to distinguish many sets of words, such as easy/easing and hammer/hanger. A two-pass process is used to recognize the unstressed syllables. First, the system segments the signal by broad phonetic classes. Then, routines are applied to identify the final segment of the word and to determine a list of some of the features of the segment preceding the final segment. Most of the decisions made by the system are based on energy measurements and formant frequencies. System testing showed that a segment in an unstressed syllable could be reliably identified as one of a small set of phonemes, but absolute identification is difficult. Broad-class segmentation was 83% accurate, and the best feature-recognition rates were for strong fricatives and voicing. Final segments were identified correctly 67% of the time. Two-thirds of the errors, however, were substitutions within pairs of phonemes which are only rarely used to distinguish words.