In the paper a method or speeding up the response of a CDHMM based speech recognition system is introduced. The method, applicable for the recognition of discrete utterances, uses a two-level classification scheme. It consists in a fast match done with simplified models, followed by a final accurate match with a limited number of selected standard models. In this way the recognition time can be reduced by great deal without any significant loss of recognition accuracy. The method has been successfully applied in the design of real-time speech recognition systems operating with small and middle-size vocabularies.
