Controlling the listener response rate of virtual agents

This paper presents a novel way of interpreting the prediction value curves that are the output of the current state-of-the-art models in predicting generic listener responses for embodied conversational agents. Based on the time since the last generated listener response, the proposed dynamic thresholding approach varies the threshold that peaks in the prediction value curve need to exceed in order to be selected as a suitable place for a listener response. The proposed formula for this dynamic threshold includes a parameter which controls the response rate of the generated behavior. This gives the designer of the listening behavior of a virtual listener the tools to adapt the behavior to the situation, targeted role or personality of the virtual agent. We show that the generated behavior is more stable under changing conditions than the behavior of the traditional fixed threshold.