In extractive spoken document summarization, it is desired to select important utterances from documents to construct the summary while avoiding redundancy among the selected utterances, but it is not easy to balance the two different goals. In this paper, a supervised spoken document summarization approach is proposed based on structured support vector machine (SVM), in which the above two goals are jointly considered during training. A set of parameters not only describing the ways to evaluate the importance of the utterances but minimizing the redundancy is directly learned from the training set. Encouraging results were obtained on a lecture corpus in the preliminary experiments.

Index Terms: speech summarization, structured SVM