Stress and vowel duration effects on syllable verification

Abstract

Recognition system for continuous speech must be able to deal with many sources of acoustic variation, such as variation in stress. We have examined the performance of a speaker-dependent syllable recognition task for a small group of stressed and unstressed syllables. These syllables, some containing the long vowel /æ/ and others the short vowel /I/, were excised from continuous speech and parameterized in terms of mel-scale cepstral coefficients at two levels of precision. From these data, four types of template dictionaries were formed: A contained individual templates for each stressed and unstressed syllable, B contained templates formed by warping stressed and unstressed syllables together, C contained only stressed syllables, and D contained only unstressed syllables. The recognition error rate increased when the template dictionaries were used in the order A, B, C, and D. The results indicate that training data should contain stressed and unstressed tokens, but for most syllables these may be combined into one template. Further results point out conditions for which the inclusion of stress is of importance in dictionary construction. [Work supported by NSF grant MCS-7916177.]