Prediction of Team Cohesion, by examining prosodic mimicry in small group meetings

In this thesis the alignment of speech behavior between team members of small group meetings (3-8 people) and its relation to team cohesion is studied. Since team cohesion is highly related to effectiveness and productivity, automatically estimating cohesion can be a useful tool for assessing the meeting quality. This information can be fed back to the meeting groups with the goal of improving meeting quality. Research within social sciences suggests that participants of a meeting put larger effort in aligning their speech behavior when they like each other more and when they are more committed to the common task. We therefore hypothesize that non-verbal speech alignment is positively related to both constructs of team cohesion, social and task cohesion. We propose a novel method of quantifying group-level alignment of non-verbal speech behavior, namely, voice intensity, pitch, Mel-frequency cepstral coefficients and speech rate. The resulting quantifications of non-verbal speech alignment are used to estimate high and low social and task cohesive regions. This is the first study examining alignment within prosodic cues on a group level, instead of a dyadic (one-to-one) level. In total, more than 20 hours of naturalistic group meetings were recorded and examined. As a measure for cohesion, a behavioral approach was followed, in which the cohesion levels were objectively quantified using verbal expressions. As cohesion is known to be inherently temporal, a dynamic methodology was pursued quantifying the alignment of speech on small time windows. By extracting three features solely related to the alignment of prosodic speech behavior, it was found that 2-minute high and low social cohesive regions could be classified with 0.72 Area under the ROC curve. The results imply that the alignment of speech behavior between the different participants in a meeting is a strong indicator for the team cohesion. This result holds for small naturalistic group meetings, implying that the proposed methodology is capable of detecting prosodic speech alignment on a group level. Predicting task cohesion was more challenging, an accuracy of 0.64 Area under the ROC curve was obtained. This performance was better than the Area under the ROC curve obtained with features related to turn-taking patterns (AUC). It appears that social cohesion is more openly expressed by behavior than task cohesion is. This insight accentuates the importance of distinguishing in social and task cohesion, instead of integrating them as one construct "team cohesion". Further analysis is needed to draw firm conclusions about the underlying behavior examined and how exactly the dynamics of the meeting are related to group outcomes as productivity and meeting satisfaction.Electrical Engineering, Mathematics and Computer ScienceComputer Vision / Pattern RecognitionSignals and Systems, Electrical Engineerin