Finding and using stress accent in spoken language understanding

Abstract

Stress accent is a feature of the prosodic system in English that functions syntagmatically to organize an utterance into shorter sections. Knowledge about the location of stress accents in the segmental stream can help a speaker‐independent spoken language understanding system prune the search space, e.g., for the correct lexical item or the correct pragmatic interpretation. Unlike automatic systems, human listeners are able to locate accents because they know what accented versus unaccented syllables sound like and because they can normalize for such things as speaker and speaking rate. Here, the results of a study designed to locate accents in continuous speech are reported. An attempt is made to use both speech knowledge and normalization by identifying salient characteristics of accent in training data along with accent characteristics of the test utterance. Accented and unaccented syllables are separated using selected clustering algorithms on a feature space of energy and duration measures. In addition, some specific ways in which the output of the accent finder can be used by other components of the system (e.g., lexical access, pragmatics) are suggested. [Work partially supported by DARPA/ISTO.]