A spoken dialogue system with active/non-active word control for CD-ROM information retrieval

Abstract

This paper describes a development of a spoken dialogue travel guidance system, TARSAN. TARSAN uses commercial CD-ROM guidebooks as its knowledge source, containing a large amount of travel information. To deal with this amount of information, a large vocabulary has to be accepted by a speech recognizer without reducing its performance. Thus, we propose two steps of active/non-active word control methods: (1) a word/grammar prediction strategy, and (2) unknown word re-evaluation algorithm. The word/grammar prediction strategy dynamically changes a recognition network according to a conversation situation by making use of results retrieved from the CD-ROMs. This strategy makes users to access almost all data on the CD-ROMs using a small vocabulary speech recognizer. The unknown word re-evaluation algorithm processes unknown words and non-active words using Garbage Models by integrating them into the recognition network, and once the Garbage Models are recognized, the unknown part will be compared with the non-active words. This algorithm enhances the ability of the word/grammar prediction. In the experiment without Garbage Models, 80.9% of the utterances were correctly understood. In the unknown word re-evaluation experiment using the Garbage Models, 86.4% were correctly re-evaluated, while the false alarms of 5% were found.