Abstract
This contribution is in the field of automatic detection of conflict in group discussions from voice analysis. A reliable detector of conflict would be useful for many applications, such as security in public places, the quality of customer services, and the deployment of intelligent agents. Experiments were conducted on the SSPNet Conflict Corpus during the Interspeech 2013 Conflict Challenge. The audio clips, which were extracted from political debates, have been classified into two classes of conflict level (low or high). In this study, we have used the turn-taking characteristics, such as interruptions, for improving the conflict detection. In a group discussion, overlapping speech (overlap) corresponds to interruption. Two overlap detectors have been developed using the SVM classifier and audio features. The first detector aims at detecting whether interruptions occur in a speech segment. The second detector aims at detecting when interruptions occur in a speech segment and whether these interruptions are related to low- or high-level conflict. A multi-expert architecture has been defined to incorporate the knowledge that arises from the interruption detectors. The two-class conflict detector (low or high conflict) consists of an SVM classifier that uses a composite feature set as input. This feature set is a concatenation of selected audio features and overlap detector-based features. Experiments provide an unweighted accuracy recall (UAR) of 85.3 % on the Test set. These results indicate an improvement of 4.5 % compared to the official baseline system results. In conclusion, the interruptions in speech can be detected and can significantly improve the automatic conflict detection.