A topic classification system based on parametric trajectory mixture models

In this paper we address the problem of topic classification of speech data. Our concern in this paper is the situation in which there is no speech or phoneme recognizer available for the domain of the speech data. In this situation the only inputs for training the system are audio speech files labeled according to the topics of interest. The process that we follow in developing the topic classifier is that of data segmentation followed by the representation of the segments by polynomial trajectory models. The clustering of acoustically similar segments enables us to train a trajectory Gaussian mixture model that is used to label segments of both on topic and off topic data and the labeled data enables us to create topic classifiers. The advantage of the approach that we are pursuing is that it is language and domain independent. We evaluated the performance of our approach with several classifiers demonstrated positive results.