Development of Method for Detecting Conversation Breakdown Using Behavioral Information

Abstract

Existing conversation systems have difficulty in communicating smoothly. One of the reasons is that the robot cannot recognize the conversation breakdown. To solve this problem, we attempt to detect conversation breakdown using human motion characteristics. According to our preliminary research, features were found in head Yaw movement when conversation breakdown occurred. As a result of FFT analysis of yaw data, the low-frequency power was different between normal conversation and conversation breakdown. Based on this finding, the energy rate, which relatively expresses the magnitude of the frequency power, was defined. The conversation breakdown detector was developed to estimate conversation breakdown when the z-test shows a difference between the energy rate of normal conversation and the energy rate of the conversation breakdown.