On not recognizing disfluencies in dialogue

This paper tests the hypothesis that listeners miss disfluencies or fail to transcribe them accurately because disfluencies interfere with the normal relationship between speech sound and linguistic context in human spoken word recognition. In a word-level gating experiment 16 listeners heard a total of 56 disfluent utterances selected from a corpus of spontaneous speech, 56 length-matched fluent controls, and 56 fluent foils. The proportion of words never recognized was greater in disfluent utterances than in controls. The failures clustered around the point where the disfluency interrupted the utterance, ocurring particularly within the reparanda, but were not found at corresponding locations in uninterrupted controls. Repetition disfluencies, where pre- and post-interruption portions might easily be construed together, allowed more successful word recognitions than recast disfluencies, where reconstruction of a single intended utterance would be difficult, if not impossible. The results have implications both for understanding human speech recognition and for improving the robustness of ASR systems.