Speaker Prediction based on Head Orientations
Rutger Rienks, Ronald W. Poppe, Mannes Poel · University of Twente Research Information · 2005
To gain insight into gaze behavior in meetings, this paper compares the results from a Naive Bayes classifier, Neural Networks and humans on speaker prediction in four-person meetings given solely the azimuth head angles. The Naive Bayes classifier scored 69.4% correctly, Neural Networks 62.3% and humans only 37.7%. None of the classifiers was able to generalize over meetings. We show that there are strong indications that human specific gaze behavior influences the fact that the models do not generalize. Additionally, we show that for all classifiers the performance of the prediction in the beginning and at the end of a speaker turn is worse than halfway through the speaker turn.