Segmentation and classification of TV news articles based on speech dictation

We propose a method to automatically segment continuous TV news speech into articles and classify them into 10 topics based on speech dictation techniques using speaker independent triphone HMMs and word bigram. The proposed method is composed of keyword selection and topic function which computes the similarity between topics and the analytical period in spoken sentences. In the keyword selection, relative mutual information is proposed and compared with other 5 measures. It showed the highest score 81.8% and 87.0% in topic classification and topic boundary detection respectively. In the topic function, we compared four methods and the normalized association showed the best scores in topic segmentation and classification.