Modeling the synchrony between interacting people: application to role recognition

Abstract

The study of social interactions has attracted increasing attentions. The role recognition is one of its possible applications and the core of this study. This article proposes some approaches to automatically recognize the role of the participants of a meeting by modeling the synchrony of temporal nonverbal audio features. In our approache the Influence Model (IM), a Hidden Markov Model (HMM)-like, is used to model this synchrony and to extract from input data a feature vector that contains both information about temporal transitions (intra-personal data) and interaction between participants (inter-personal data). This modeling of the meeting is used as input of a Random Forests (RFs) for the role recognition task. The experiments are performed on 138 meetings (approximately 45 hours of recordings) from Augmented Multiparty Interaction (AMI) Corpus. Accuracy scores show that this combination of generative (IM) and discriminative (RFs) approaches permits to outperform state-of-the-art role recognition rates.