Sushanth Dabbiru

Statistical Modeling for Story Segmentation of Audio Broadcasts

Date: November 5, 2004

Abstract

Recent advances in data storage techniques have made large amounts of audio and video data from multimedia systems easily accessible. However, the problem of accessing such large amounts of data has caused a significant improvement in data storage and management using advanced audio and video management systems. A lot of these systems have recently begun to value speech as an archival source of information. The major problem with speech is that its transcribed version is raw text data, which has to be converted to a structured format before any information retrieval is to take place. Story segmentation is one of the many techniques used to create structured documents, making it essential in applications like audio indexing, data mining and information retrieval. Our work is focused on segmenting speech transcriptions from English and Arabic broadcast news into stories. Story Segmentation imposes a document structure on transcribed audio data from broadcast news source, enabling the user to find relevant information. The story segmentation system developed during the course of this thesis uses a statistical model to detect story boundaries in speech transcripts. In our approach, lexical-based features are used in a statistical model to hypothesize story boundaries between sentences in speech transcriptions from broadcast news. These features are discriminative since they encode the changes in the story content within raw streams of texts.

Our story segmentation system turned out to be independent of language models based on the results obtained while testing on Arabic and English. The story segmentation system has given a reasonable performance, which was based on achieving accurate scoring of story boundaries in speech transcriptions of Arabic and English. The stories produced by the story segmentation system can be used by other advanced text-based approaches for information retrieval and content management.

Committee:
Prof. John Makhoul
Prof. Vinay Ingle
Prof. Jennifer Dy