Estimating Speaker’s Engagement from Non-verbal Features Based on an Active Listening Corpus

Abstract

The elderly who live alone are increasing rapidly in these years. For their mental health, maintaining their social life with others is reported useful. Our project aims to develop a listener agent who can engage active listening dialog with the elderly users. Active listening is a communication technique that the listener listens to the speaker carefully and attentively. The listener also ask questions for confirming or showing his/her concern about what the speaker said. For this task, it is essential for the agent to evaluate the user’s engagement level (or the attitude) in the conversation. In this paper, we explored an automatic estimation method based on empirical results. An active listening conversation experiment with human-human participants was conducted for corpus collection. The speakers’ engagement attitude in the corpus was subjectively evaluated by human evaluators. Support vector regression models dedicated to the periods when the speaker is speaking, the listener is speaking, and no one is speaking are built with non-verbal features extracted from facial expressions, head movements, prosody and speech turns. The resulted accuracy was not high but showed the potential of the proposed method.