Aiming at the problem of underdetermined blind separation of the attenuated and delayed mixtures

we propose a method based on the subspace decomposition to find the single source area in time-frequency domain

and then estimate the mixing matrix via clustering the principal eigenvectors of the covariance matrixes corresponding to the single source regions without knowing the number of original sources

finally we modify the algorithm based on subspace projection to extract the original signal.Simulation results indicate that the proposed algorithm estimates the mixing matrix with higher accuracy and separates the source signals with higher gain

compared to the other algorithms