Errors, Repetition, and Contrastive Emphasis in Speech Recognition

In repeating an utterance for the benefit of a listener, talkers can use contrastive stress to mark those parts of an utterance that were misrecognized. In this study, we found that talkers use a consistent set of markers to indicate contrastive stress: Listeners unaware of the nature of the misrecognition can readily identify a contrastively stressed word. Based on an analysis of these markers, we have implemented an automatic algorithm that identifies contrastive stress with about the same accuracy as humans, using amplitude, duration, silence, and pitch cues. This ability to detect contrastive stress can be effectively exploited, as part of error recovery strategies, by a recognition system.