Designing for errors: similarities and differences of disfluency rates and prosodic characteristics across domains

This paper focuses on some characteristics of disfluencies in human-human (HHI) and human-computer (HCI) interaction corpora to outline similarities and differences. The main variables studied are disfluency rates and prosodic features. Structured, table-like input increases the disfluency rate in HCI and decreases it in HHI. Direct exposure (visibility) to the interface also increases the rate and gives speech a unique prosodic pattern of hyperarticulation. In most of the studied corpora, silences at the disfluency site are not predicted by syntactic rules. Similarities between HCI and HHI exist mainly in the prosodic realizations of the reparandum and the repair. The findings contribute to better understanding and modeling of disfluencies. Speech-based interfaces need to focus on communication types that are well-understood and prone to good modeling.