A solution is proposed to deal with the problem that ″less number of features cannot coexist with higher recognition rate″ in the traditional theory of speaker recognition. Ladder observation matrix projection is used to change the traditional Mel-frequency cepstral coefficient (MFCC) parameters based on compressed sensing theory, presenting a new recognition parameters named compressed sensing MFCC (CS-MFCC) parameters. These parameters make storage capacity decrease to less than 1/n of the original, here n is the compression ratio of the line ladder matrix, and also greatly increase the robustness of the system. Furthermore simulation results prove that when n is 4, the recognition rate increases to 96% above.