Modeling Vocal Interaction for Text-Independent Classification of Conversation Type
Kornel Laskowski, Mari Ostendorf, Tanja Schultz · 2007
We describe a system for conversation type classification which relies exclusively on multi-participant vocal activity patterns.Using a variation on a well-studied model from stochastic dynamics, we extract features which represent the transition probabilities that characterize the evolution of participant interaction.We also show how vocal interaction can be modeled between specific participant pairs.We apply the proposed system to the task of classifying meeting types in a large multi-party meeting corpus, and achieve a three-way classification accuracy of 84%.This represents a relative error reduction of more than 50% over a baseline which uses only individual speaker times (i.e.no interaction dynamics).Random guessing on this data yields an accuracy of 43%.