A continuous model for the management of turn-taking in user-agent spoken interactions based on the variations of prosodic signals

Abstract

Many recent works on agent architectures have focused on polite and optimal turn-transitions. However, real turn-taking is more complex due to several contextual variables, linked to each agent’s own goals (cooperative or non cooperative for example). For mixed-initiative interactions, we need to go beyond the polite agent context to make more complex patterns of turn-taking emerge. We present here an architecture based on a dynamical and continuous model of turn-taking, able to control the turn-taking behaviors of the agent depending on its willingness to speak or not. We show how we implemented our model based on human data and how complex patterns of turn-taking emerge from agent-agent simulations. Finally, we present the results of a perceptual experiment where we questioned participants about the intentions of two agents interacting.