For many practical speech recognition applications, the rejection of out-of-vocabulary(OOV) words is an important issue. To make the rejection decision, confidence measures are computed in many systems. In this paper we focus on the problem of isolated word rejection and our strategy is based on a postclassifier. First of all, by using a linear classifier, we combine several promising features presented by others to obtain a confidence measure. By comparing the confidence value with the decision threshold, we can reject the OOV words. After that we present two novel features which are proved effective in our experiment. Finally, because many data sets are linearly nonseparable, we present a framework based on Support Vector Machine(SVM). The experiments show that a considerable improvement(about 74%) of the equal-error-rate(EER) is achieved after all improvement strategies are integrated into the system.