Turn-taking patterns in human discourse and their impact on group communication service design

Recent studies demonstrated the benefit of integrating speaker prediction features into the design of group-communication services supporting multiparty online discourse. This paper aims at delivering a more elaborate analysis of speaker prediction by analyzing a larger volume of data. Moreover, it tests the existence of speakers dominating speaking time. Towards this end, we analyze tens of hours of recorded meeting and lecture sessions. Our principal results for meeting-like interaction manifest that the next speaker is one of the last four speakers with over 90% probability. This is seen consistently across our data with little variance (standard deviation of 8.71%) independent of the total number of potential speakers. Furthermore, lecture time is in most cases significantly dominated by the tutor. In meetings, although a single dominating speaker is always evident, domination exhibited high variability. Generally, our findings strengthen and further motivate the act of incorporating user-behavior awareness into group communication service design.