Modeling vocal interaction for textindependent detection of involvement hotspots in multi-party meetings

We describe a system for conversation type classification which relies exclusively on multi-participant vocal activity patterns. Us-ing a variation on a well-studied model from stochastic dynamics, we extract fea-tures which represent the transition proba-bilities that characterize the evolution of par-ticipant interaction. We also show how vo-cal interaction can be modeled between spe-cific participant pairs. We apply the pro-posed system to the task of classifying meet-ing types in a large multi-party meeting cor-pus, and achieve a three-way classification accuracy of 84%. This represents a rela-tive error reduction of more than 50 % over a baseline which uses only individual speaker times (i.e. no interaction dynamics). Ran-dom guessing on this data yields an accuracy of 43%.