Abstract: In blind source separation, to estimate precisely the mixing matrix is a challenging problem, especially with more original source than sensors (i.e., underdetermined case).Most existing algorithms estimate the mixing matrix by using the sparsity of signals in the time-frequency (TF) plane and the assumption that the time-frequency distributions of the input sources do not overlap.However,estimation performance is shown to be limited by the sparsity of signal, especially when the time-frequency distributions of the input sources overlap each other much. In this paper, we proposed a novel algorithm for mixing matrix estimation with less sparsity of the input signal. First, we use the proposed algorithm identifies the single-source-points by comparing the absolute ratio value of the real and imaginary parts of the TF transform coefficient of the mixed signal. Second, the column vectors of mixing matrix will be estimated by clustering algorithm. The proposed algorithm makes a good performance for estimating the mixing matrix and it is simpler than existing algorithm.