Itegrating backchannel prediction models into embodied conversational agents

In this paper we will present our design for generating listening behavior for embodied conversational agents. It uses a corpus based prediction model to predict the timing of backchannels. The design of the system iterates on a previous design (Huang et al. [5]) on which we propose improvements in terms of robustness and personalization. For robustness we propose a variable threshold determined at run-time to regulate the amount of backchannels being produced by the system. For personalization we propose a character specication interface where the typical type of head nods to be displayed by the agent can be specied and ways to generate slight variations during runtime.