This paper presents a new algorithm to identify matrix knowing only their multiplication . Where is sparse and . The data used for matrix identification are chosen by Least Square method, whose fitting errors are smaller than a given threshold. Then, K-means clustering method is adopted. This technique avoids data overlapping at the origin, thus improving the accuracy of mixing matrix estimation. The validity of the method for true voice separation is verified by computer simulation. Also comparison with other methods is made to verify the efficiency of the algorithm. Simulations show that the algorithm has the property of accuracy and low-cost computation.