This paper deals with the problems of single word spotting and non vocabulary word rejection, with the purpose of enhancing recognition performance of conventional isolated word recognizers. The approach proposed in this paper is organized in three steps: 1) using a garbage model to represent extraneous speech, garbage model is generic HMM with a tied covariance matrix and trained on a wide range of speech with different characteristics. 2) applying a grammar-driven frame synchronous level building algorithm to generate three possible combinations of key word and extraneous speech, the patterns of the three candidates for un utterance are: only a key word or no key word at all, a key word with extraneous speech at the beginning or end; a key words with extraneous speech at both the beginning and end. 3) adding a post decoder to perform a sequence of validity tests on the multiple candidate strings. The candidate that passes all testing and still holds the best likelihood score is selected as spotted word.