Patterns of gaze in speech agent interaction

Abstract

While gaze is an important part of human to human interaction, it has been neglected in the design of conversational agents. In this paper, we report on our experiments with adding gaze to a conventional speech agent system. Tama is a speech agent that makes use of users' gaze to initiate a query, rather than a wake word or phrase. In this paper, we analyse the patterns of detected gaze when interacting with the device. We use k-means clustering of the log data from ten users tested in a dual-participant discussion tasks. These patterns are verified and explained through close analysis of the video data of the trials. We present similarities of patterns between conditions both when querying the agent and listening to the answers. We also present the analysis of patterns detected when only in the gaze condition. Users can take advantage of their understanding of gaze in conversation to interact with a gaze-enabled agent but are also able to fluently adjust their use of gaze to interact with the technology successfully. Our results point to some patterns of interaction which can be used as a starting point to build gaze-awareness into voice-user interfaces.