Support vector machines for discourse segmentation in meetings

We build a system for the automatic detection of topic boundaries in meeting recordings. We extract a number of prosodic and lexical features from a single tabletop microphone and the human-generated transcript. We use a support vector machine trained on these features to learn the properties of topic boundaries. In a detection experiment, we see results with up to 98% correct accept rate with a 5% false alarm rate, a significant gain in performance over simple modeling techniques, such as single-mixture Gaussian mixture models.