Topic Detection in Read Documents

In this paper, we address the importance and the problems involved in topic annotation in the speech retrieval domain. Identified the problem, an algorithm developed to perform automatic topic annotation of broadcast news (BN) speech corpora is described. The approach adopted is based in Hidden Markov Models (HMM) and topic language models, to solve topic segmentation and labelling tasks simultaneously. To overcome the lack of topic labelled material to train the statistical models, a two-stage unsupervised clustering was developed. Both stages are based on the nearest-neighbour search method, using the Kullback-Leibler as a distance measure. On-going experiments to evaluate the system performance are also described.