Voice based interface for route search system

Abstract

In this paper we are presenting our approach of creating voice based interface for one of the leading Lithuania bus route search system — www.autobusubilietai.lt. We designed a hybrid speech recognition system which is based on one Lithuanian speech recognizer (LIEPA) and two foreign language recognizers (German and Spanish). We experimented with different methods that may be used for combining outputs of these recognizers: k-Nearest Neighbors, Naïve Bayes, Support Vector Machines, Decision tree (CART), Linear discriminant analysis and Logistic regression. We achieved 88.0 % recognition accuracy by using 3-Nearest Neighbors as combination method. Compared with best single recognizer we increased recognition accuracy by 3.5 %. Our experiments proved that recognition accuracy can be improved by combining outputs of different speech recognizers. What is more by using voice based interface we decreased time spent for route search from 10–15 sec. to 5–10 sec. and improved end user satisfaction.