Gaze and filled pause detection for smooth human-robot conversations

Abstract

Let the human speak! Interactive robots and voice interfaces such as Pepper, Amazon Alexa, and OK Google are becoming more and more popular, allowing for more natural interaction compared to screens or keyboards. One issue with voice interfaces is that they tend to require a “robotic” flow of human speech. Humans must be careful to not produce disfluencies, such as hesitations or extended pauses between words. If they do, the agent may assume that the human has finished their speech turn, and interrupts them mid-thought. Interactive robots often rely on the same limited dialogue technology built for speech interfaces. Yet humanoid robots have the potential to also use their vision systems to determine when the human has finished their speaking turn. In this paper, we introduce HOMAGE (Human-rObot Multimodal Audio and Gaze End-of-turn), a multimodal turntaking system for conversational humanoid robots. We created a dataset of humans spontaneously hesitating when responding to a robot's open-ended questions such as, “What was your favorite moment this year?”. Our analyses found that users produced both auditory filled pauses such as “uhhh”, as well as gaze away from the robot to keep their speaking turn. We then trained a machine learning system to detect the auditory filled pauses and integrated it along with gaze into the Pepper humanoid robot's real-time dialog system. Experiments with 28 naive users revealed that adding auditory filled pause detection and gaze tracking significantly reduced robot interruptions. Furthermore, user turns were 2.1 times longer (without repetitions), suggesting that this strategy allows humans to express themselves more, toward less time pressure and better robot listeners.