This paper focuses on the mixing matrix estimation in underdetermined blind source separation.A method for detecting the single source points in the time-frequency (TF) domain is given to reduce the effect of multiple source points in the TF domain. A method of outlier detection is used to remove outliers
in the single source points to improve accuracy of mixing matrix estimation. A subtractive clustering method is used to cluster the single source points so that the number of sources and the mixing matrix are obtained simultaneously. Experiments on speech signals show that the proposed algorithm can estimate the mixing matrix with high accuracy and robustness as compared to the other two mixing matrix estimation algorithms based on single source point detection in the TF domain.