Abstract: The assumption of W-disjoint orthogonality(W-DO) can simplify the problem of blind separation for under-determined mixed speech signals at the cost of decreasing the separation performance.A method based on principal component analysis(PCA) is proposed to detect the time-frequency cells where only one source exists,under the as-sumption of approximate W-DO of speech signal.All the detected time frequency cells satisfy the W-DO,so that the estimation precision of the mixing matrix is improved.The direction of arrival(DOA) of sources is estimated from the mixing matrix and is exploited to solve the permutation ambiguity problem.Simulation results demonstrate that the proposed method outperforms the typical DUET method,the average signal to interference(SIR) is improved by 2.77dB.