A CORPUS-BASED STUDY OF REPAIR CUES IN SPONTANEOUS SPEECH

The occurrence of disfluencies in fully natural speech poses difficult challenges for spoken language understanding systems. For example, al though self-repairs occur in about 10% of spontaneous utterances, they are often unmodeled in speech recognition systems. This is partly due to the fact that little is known about the extent to which cues in th e speech signal may facilitate automatic repair processing. In this pa per, acoustic and prosodic cues to self-repairs are identified, based on an analysis of a corpus taken from the ARPA Air Travel Information System database, and methods are proposed for exploiting these cues fo r repair detection, especially the task of modeling word fragments, an d repair correction. The relative contributions of these speech-based cues, as well as other text-based repair cues, are examined in a stati stical model of repair site detection that achieves a precision rate o f 91% and recall of 86% on a prosodically labeled corpus of repair utt erances.