This paper wishes to contribute to the solution of the problem occurring when an automatic speech recognition module does not recognize an input utterance correctly and delivers wrong words to an understanding block and to a dialog manager, which consequently causes a false dialog system response to a user. The solution is based on an introduction of a recognition confidence measure, which evaluates the belief that the recognition result is accurate. In our previous paper [4] the confidence measure was based on a difference between so-called a mumble model [2,4] and a recognition network scores, and on a comparison of this score difference to a heuristically set threshold. In this paper we present a new statistically based technique which uses statistical models and Bayes decision rule. The paper also briefly describes the structure and behavior of the mumble model and its implementation in a real-time telephone dialog system. The main part of this article deals with a feature selection for classification and acceptance/rejection statistical model parameters estimation. The new rejection technique has been evaluated on the Czech yellow-pages database. Experimental results show that the proposed rejection technique achieves approximately 12% equal error rate (EER) in comparison with 20% EER of the previously one, which represents 63% relative EER improvement.