In this report we describe detection of pronunciation words which are not included in the vocabulary of speech recognition system, using method confirmation correctness of recognition for separate phones. The algorithm of speech recognition system work is described. Formulas for calculation distances between sound models, used for definition the model nearest to a recognized sound, and also the model most far from it, are given. The results of experiments, confirming serviceability of this method, are described. Disadvantages of method and direction of its further development are discussed.