When can listeners detect disfluency in spontaneous speech?

Three experiments investigated listeners' ability to detect disfluency in spontaneous speech. All employed gated word recognition with judgments of disfluency for spontaneous utterances containing disfluencies and for three kinds of fluent control utterances from the same six speakers: repetitions of corrected recordings of original disfluent items, spontaneous fluent utterances loosely matched in structure to the disfluent items, and repetitions of those spontaneous fluent items. In Experiment 1, 120 stimuli were word-level gated and presented to 20 subjects for word identification and for judgments on whether the utterance was about to become disfluent. Listeners were unable to predict disfluency reliably. New subjects (N = 20, 43) judged whether the same utterances had already become disfluent at each word gate in Experiment 2 or at each 35 ms gate in Experiment 3. Subjects reliably detected existing disfluencies during the first word gate after the interruption and before they recognized the word. Though more common around disfluencies than at similar points in controls, failures of word identification were not reliably associated with detection. Results are discussed in the light of computational models of disfluency detection.