Abstract

Currently, many research is being done regarding the development of Automatic Speech Recognition (ASR) systems. The ASR systems currently available can be seen as 'slot filling' machines. They pose directed questions to users who subsequently have to answer these questions in a very restricted, directed way in order to make speech recognition possible. The mere recognition of speech using neural networks is the main issue concerning the development of this kind of systems. However, when an ASR system has to handle free, non-restricted speech, the identification of the meaning of the words spoken to the system becomes the most important issue.

By order of the Dutch'Openbaar Vervoer Reisinformatie' (O.V.R.) corporation the Technical University of Delft is making a feasibility study of the possible application of an ASR system, with the help of which people can obtain traveling information by calling to the O.V.R. telephone service (06-9292).

A 'slot-filler ASR system' is already available to the O.V.R. As explained before such a system asks directed questions to the caller (e.g. 'from which place do you want to depart?). If the callers answer to a question is unclear to the system, it simply re-asks the same question. This process continues until the caller does give a clear answer.

In the framework of the feasibility study done by the TU-Delft, research is being done towards the application of an ASR system being able to handle free speech. As said before, in order to make this possible the meanings of the words spoken by a caller have to be understood.

This thesis can be seen as a starting point in the design of an 'O.V.R. ASR system', being able to handle free speech. The main issue in the development of such a system is the design of a 'higher level, conceptual dialogue structure', with the help of which the concepts (the meanings) of utterances can be expressed. In this thesis a description will be given of the design of such a structure, which has been designed corpus-based. This means that the dialogues carried out in reality on the 06-9292 information service, serve as the basis for the development of the conceptual structure.

Using the results of the conceptualizations of the dialogues, the dialogues have been clustered into a couple of groups of different complexity, using a statistical technique called factor analysis. Furthermore, in order to be able to say something about the follow-up probabilities of the conceptualizations, other statistical analyses have been made. Follow-up probabilities have been calculated based on bi-class and bi-gram methods.

The most important component of any ASR system is the so-called dialogue manger. Using a dialogue manager an ASR system identifies meanings of utterances, calculates which utterances are most likely to follow and keeps track of the 'plans' users have. Using the results of the conceptualizations of the dialogues and the statistical analyses a start has been made in the design of a dialogue manager.

To test the clarity, unambiguity and completeness of the conceptual structure an application called 'the Generic Dialogue Tool' has been designed. Using this tool the impact of the information which is lost as a result of the datareduction (dialogue to a higher level conceptual structure) has also been investigated.

Finally a description will be given of the design of an application called the "Train Information Pillar'. With the help of this system people can obtain information about train journeys by pushing buttons. The information is given to the user in text-output or speech-output form, which can be chosen by the user. The buttons of the system in fact stand for the concepts of the utterances as present in the O.V.R. dialogues.