Track:
Contents
Downloads:
Abstract:
Continuous speech recognition technology has recently matured to the point where it has become feasible to develop spoken dialogue interfaces to computer systems that help people perform tasks in simple domains. Word recognition accuracy for spontaneous dialogue, however, is still a long way from perfect, and any system that uses spoken natural language input has to have some mechanism for dealing with speech recognition errors. In fact, since we cannot expect speech recognition technology ever to surpass human performance, which is itself imperfect, the handling of misrecognition will continue to be an issue that needs to be addressed by spoken dialogue systems. Many techniques have been incorporated into various components of spoken dialogue systems to try to provide robustness in the presence of speech recognition errors. This paper describes a study that explores how humans go about signaling communication problems caused by speech recognition errors. The study is the first step in designing a conversational agent that has human-like repair initiation and feedback behavior, and can perhaps provide robustness by resolving communication problems within the course of the dialogue itself.