3Department of Computer Engineering and Information Technology, Razi University, Kermanshah, IRAN
One of the most important steps in a keyword spotting (KWS) system is a post-processing procedure to compute a confidence measure (CM) for each hypothesized keyword. The CM is commonly estimated by likelihood-based acoustic scores. However durations of the detected keyword, which include useful information, has not been studied directly in the KWS systems. In this paper, three duration-based features are proposed for such system. Also, using linear discriminant analysis (LDA), the proposed duration-based features are discriminatively combined with the acoustic scores and used as the final discriminative CM (DCM) in the KWS system. The proposed DCM results up to 11.3% improvement in FOM against to the conventional likelihood-based CM over a Persian conversational telephone speech database.