Helping, I Mean Assessing Psychiatric Communication: An Applicaton of Incremental Self-Repair Detection

Abstract

Self-repair is pervasive in dialogue, and models thereof have long been a focus of research, particularly for disfluency detection in speech recognition and spoken dialogue systems. However, the generality of such models across domains has received little attention. In this paper we investigate the application of an automatic incremental self-repair detection system, STIR, developed on the Switchboard corpus of telephone speech, to a new domain – psychiatric consultations. We find that word-level accuracy is reduced markedly by the differences in annotation schemes and transcription conventions between corpora, which has implications for the generalisability of all repair detection systems. However, overall rates of repair are detected accurately, promising a useful resource for clinical dialogue studies.