Abstract: Using the data mining techniques to recognize the speech sources, certify the speaker identities and assign the operation permissions is quite meaningful in both theoretical and practical senses. This paper mainly investigates two types of speech recognition. One is based on the same voice contents, while the other is on different voice contents. For the algorithms, the widely used Mel frequency cepstral coefficient (MFCC) algorithm is adopted for the feature extraction; and dynamic time warping algorithm are combined to classify the patterns. In particular, K-means₊₊ algorithm and principle component analysis algorithm are added before the use of dynamic time warping algorithm for the second type. As a result, in the type of the same voice contents, once an appropriate threshold is selected, a good recognition effect can be derived.