Panelist: David DeVault
Title: Toward Fluid Conversational Interaction in Spoken Dialogue Systems

Abstract:

Human-human conversation is highly interactive, on a very rapid
time-scale, with participants providing a range of overlapping and
low-latency responses and signals, as speech is occurring, to
facilitate a rapid and successful communication process.  My research
aims to advance our understanding of how such fluid turn-taking and
interaction can be modeled and replicated computationally.  My work
connects to the themes of this workshop in several areas.  One is in
modeling the real-time flow of information and communicative
functions in conversation.  I see opportunities in this area to use
empirical methods to improve how dialogue systems model the
communication channel with their users.  In particular, achieving
fluid conversation with machines requires that we move away from
simplistic turn-taking protocols and unrealistic assumptions about how
conversation is structured in time.  Instead, we need to create
fine-grained models that take seriously the timing of speech and the
detailed mechanics of how human conversation unfolds through
interlocutor decision-making at small timescales.  A second area is in
understanding the value of when to intervene vs. staying silent and
saying nothing.  For dialogue systems as a field, we need to improve
our ability to create intelligent agents and other dialogue systems
that are not purely user-driven or system-driven, but instead have a
pervasive mixed-initiative capability.  But doing this well requires,
among other things, deeper and better models of the right time for
systems to speak vs. wait.  I will discuss recent research at USC that
investigates this decision-making in a couple of implemented spoken
dialogue systems.  There are additional connections to methods for
scaling up dialogue systems beyond simple closed domains and the need
for achieving expressive text-to-speech synthesis with human-like
prosody.