Communication error detection using facial expressions

Abstract

Automatic detection of communication errors in conversational systems typically rely only on acoustic cues. However, perceptual studies have indicated that speakers do exhibit visual communication error cues passively during the system's conversational turn. In this thesis, we introduce novel algorithms for face and body gesture recognition and present the first automatic system for detecting communication errors using facial expressions during the system's turn. This is useful as it detects communication problems before the user speaks a reply. To detect communication problems accurately and efficiently we develop novel extensions to hidden-state discriminative methods. We also present results that show when human subjects become aware that the conversational system is capable of receiving visual input, they become more communicative visually yet naturally.