Confidence Measure Improvement Using Useful Predictor Features and Support Vector Machines

In traditional keyword spotting (KWS) systems,confidence measure (CM) of each keyword is computed from normalized acoustic likelihoods. In addition to likelihood basedscores, some keyword dependent features named predictor features such as duration and prosodic features could be definedto improve the performance of CM. In this paper a discriminative and probabilistic computation of CM based upon some useful predictor features and support vector machines(SVM) is presented for Persian conversational telephone speech KWS. Our experimental results show that higher performancewill be achieved by appending utilized predictor features. The proposed CM with linear kernel function of SVM is obtained an improvement about 8.6% in Figure-of-Merit (FOM) of KWS system

Keywords:

keyword spotting , speech recognition , confidence measure , predictor feature , SVM