Design of screen-based speaker-independent dialog system for decision support system

Abstract

The implementation of speaker-independent voice recognition system for a large volume of vocabulary is inherently difficult, or may be impossible, unless some additional information besides accoustic data is used. To support the voice communication for the Decision Support Systems (DSS) and/or Expert Systems (ES) there need to cover the selection of menu, request for "help" and the explanation of terminologies. A good news however, is that most of the voice communication can be made using the terms displayed on the screen. To support the voice dialog of these situation, this thesis proposes a scheme which combines the data displayed on the screen, accoustic data and other informations about user characteristics or profiles. This approach is feasible because the displayed items on the screen can reduce the relevant candidates to recognize to the level of speaker independence. And also, user characteristic attributes such as sex, dialect and age, etc., can be used to split the large speech base so that unnecessary candidates or comparison should be reduced. We designed the new architecture using these and automatic learning concepts for adaptation to specific user, tested the effectiveness of these user characteristic attributes in voice recognition by simulated experiments and evaluated the performance of our system with comparison to the previous research.