Interactive dialogue telephone service

Abstract

This paper considers an interactive dialogue telephone service which provides information about cultural events in Belgrade for the following week. The dialogue user-system is based on the keyword recognition in continuous Serbian. The main part of the system is dialogue manager, which has to coordinate the system components. We consider the system question and the user answer as one dialogue level. At each level, the system has to recognize a limited number of keywords. The keyword models are continuous hidden Markov models with three states for each keyword's syllable. The decoding phase confidence measure, computed on the pronounced sentence, shows the keyword has been pronounced. Thresholds and optimal step sizes for confidence measures for each keyword are determined in the training phase.