In this thesis work, we make on effort to improve the performance of the text-dependent speaker recognition algorithm based on dynamic time warping(DTW). The work is done to obtain the improved performance in the three major areas of speaker recognition problems. They how to properly extract speaker's information, how to reject imposters, and how to reduce the effects of the time-variation of the recognition rate. In order to reduce the time-variation of the matching performance of DTW, we adopt the pattern averaging technique which results in an average pattern to represent many reference patterns of long duration. We can also significantly reduce the computation time of DTW for pattern matching. To maximize the inter-speaker distinction, we propose a weighted cepstrum method. This method derives the weighted cepstral coefficients using F-ratio which is the measure of the inter-speaker variation with respect to the intra-speaker variation of the cepstral coefficients. Speaker recognition experiments are done by applying the pattern averaging and the cepstrum weighting by the F-ratio to the mel-cepstrum coefficients. The experimental results show that the speaker recognition performance improves 5 to 6 % over the conventional DTW pattern matching with cepstrum features. The proposed speaker recognition algorithm is implemented for the door-lock system using a DSP board with a TMS320C31.