Abstract

Thanks to the continuing progress in the Automatic Speech Recognition of Broadcast News (BN) audio streams it’s possible to apply Information Retrieval techniques to the transcribed text, full of recognition errors. A Topic Segmentation task is necessary for this kind of systems to perform well. In this paper a new adapted solution approach, the BITS approach, is described for segmenting continuous BN streams into homogeneous topic/story segments inside Philips’ Spoken Broadcast News Retrieval demonstrator system. This approach is based on Marti Hearst’s TextTiling approach. Some changes and improvements are made to overcome the weaknesses of the TextTiling approach working in the spoken text or transcription domain. The implemented prototype is tested and the test results will be reported in this paper.