The Application of Classical Informa - tion Retrieval Techniques to Spoken Documents

The research presented in this thesis addresses the topic of ad hoc retrieval of information from collections of spoken items such as radio news bulletins. Modern digital computers are becoming increasingly adept at processing non{ textual data, such as speech. Consequently, new methods are required to allow users to pin{point speciic items of interest in large data collections. Such a method might exploit the Hidden Markov Model (HMM), which has proved successful as the basis for many experimental speech recognition systems, and the well{understood techniques of document retrieval that have arisen from many years' research into textual information retrieval (IR). However, so far there has been little exploration of the potential combination of these methods in order to index \spoken word" data. In the IR community, several papers have put forward an approach to the problem but this approach has not been properly tested. Work done in the speech recognition area has tended to concentrate on developing systems for topic classiication. These systems are extensively pre{ trained for the task of partitioning a set of spoken messages into a set of disjoint and exhaustive classes, each one representing some topic. Their utility is, in practice, limited by the xed class set and slow operation, and they do not represent an approach to the problem of retrieving items that correspond to arbitrary topics. This thesis describes experiments combining the techniques of classical information retrieval with HMM{based speech recognition methods in order to retrieve items from a collection of spoken messages corresponding to items of radio news. In a base-line system, a new technique for wordspotting allows items matching an arbitrary expression of the information requirement to be retrieved quickly and reasonably accurately. The system is subsequently improved through the addition of appropriate language models and the use of state{of{the{art acoustic modelling. Finally, performance is compared with that obtained by two alternative approaches, including one recently proposed in the IR literature, and found to be considerably superior. 1 Declaration This thesis describes research carried out in the Speech, Vision and Robotics Group of the University of Cambridge Engineering Department between October 1991 and February 1995. It is the result of my own work and includes nothing which is the outcome of work done in collaboration. The length of this thesis, excluding references and gure captions, is forty{four thousand words. Acknowledgements Firstly, I wish to express my gratitude to Prof. Steve Young, my …