Switching Wizard of Oz for the online evaluation of backchannel behavior

Abstract

The Switching Wizard of Oz (SWOZ) is a setup to evaluate human behavior synthesis algorithms in online face-to-face interactions. Conversational partners are represented to each other as virtual agents, whose animated behavior is either based on a synthesis algorithm, or driven by the actual behavior of the conversational partner. Human and algorithm have the same expression capabilities. The source is switched at random intervals, which means that the algorithm’s behavior can only be identified when it deviates from what is regarded as appropriate. The SWOZ approach is especially suitable for the controlled evaluation of synthesis algorithms that consider a limited set of behaviors. We evaluate a backchannel synthesis algorithm for speaker–listener dialogs using an asymmetric version of the framework. Human speakers talk to virtual listeners, that are either controlled by human listeners or by an algorithm. Speakers indicate when they feel they are no longer talking to a human listener. Analysis of these responses reveals patterns of inappropriate behavior in terms of quantity and timing of backchannels. These insights can be used to improve synthesis algorithms.