Face-to-face conversation in a typical task oriented domain may be viewed
as a role play where one is having the speaker role and the other is
having the listener role. Being an attentive speaker involves many things,
but first of all creating opportunities for the listener to give
responses, such as “yeah”, “right”, “mhm”, head-nods or other gestures.
Many conversational systems operate within the “Walkie-Talkie Turn Taking
Paradigm”, i.e. one participant talks while the other is silent, and long
response times are not uncommon. To overcome this, an analysis of speech
in overlap leads to a cascade of classifiers which can advise the dialog
manager whether it is appropriate to speak in overlap or not. It is also
preferable to change ongoing behavior according to new circumstances. This
requires simultaneous perception/interpretation and production of
communicative behaviors. This seminar discusses how to implement these
considerations as an upgrade of the Embodied Conversational Agent
developed in the SEMAINE project. Finally, a pilot study is presented.