Using prosodic cues to identify dialogue acts: methodological challenge

Abstract

AbstractThis paper addresses the role of prosodic and phonetic features in talk, focussing on how to identify competitive overlaps in conversations. It has previously been claimed that turn competition may be distinguishable from non-competitive overlaps, based on phonetic/prosodic features alone. We test this hypothesis on recordings from a UK-based call center, combining a conversation analytic approach with quantitative methods using a coding scheme. Our long-term aim is to develop large-scale methods operationalizing what we know about conversational sequence and social actions into speech technological applications.Our findings show that, although there is a tendency for competitive overlaps to be more prominent in terms of loudness and pitch features than non-competitive overlaps, this difference is not sufficient to reliably identify turn competitions from the speech signal itself. We discuss the findings in relation to observations made in individual examples, and conclude by highlighting some of the methodological challenges of applying findings from the linguistic and conversation analytic literature to speech technologies.