This paper designs a DTWbased speech recognition system.The method applied in this paper belongs to the small glossarys isolated words speech recognition,which includes starting & ending point measuring,feature extraction and mode matching.The system takes preprocess of the speech signal,and then adopts the MFCC as a characteristic parameter drawing algorithm,and takes the DTW as the recognition algorithm,uses an improved template training technique to extracts the reference template with only a small quantity of samples by computing the average length.An experiment with speech signal recorded in labenvironment is given to simulate the proposed speech recognition method.The simulation results show that compared with using a single reference template,this method improves the recognition accuracy and the robustness.Furthermore,compared with the VQ,our method needs fewer samples and reduces the complexity of the algorithm.