Modelling Sub-Utterance Phenomena in Spoken Dialogue Systems

Dialogue systems that process user contributions only in chunks bounded by silence miss opportunities for producing helpful signals, such as backchannel utterances concurrent with the ongoing utterance— to the detriment of interaction quality. We survey in some detail what we call ‘subutterance’ phenomena, which such an approach misses, and then discuss two approaches to overcoming this limitation. The first is to add a ‘reactive layer’ to an otherwise unchanged, utterance-based dialogue manager. The other approach, less often taken, is to make the dialogue manager itself capable of producing such reactions. To explore the viability of such an approach, we sketch a dialogue management strategy capable of working at a sub-utterance level.