Double microphone blind source separation based on improved joint diagonalization

Abstract

Two improved methods are proposed for the problem of the underdetermined double microphone blind source separation of speech signals. Firstly, for the blind separation problem, the objective function maximization problem is equivalent to a joint diagonalization problem of a feature matrix that are simultaneously satisfied by a second-order delay correlation matrix and a fourth-order cumulant matrix. The joint approximate diagonalization algorithm of eigen matrix (JADE) has been improved to improve the separation accuracy. Further, for the underdetermined problem, that is, the number of microphones is less than the number of source signals, the improved joint diagonalization algorithm and the time-frequency (T-F) mask are combined to realize the underdetermined blind source separation based on the double microphones, and a variety of algorithms are compared with it for performance evaluation. The simulation results verify that the algorithm can accurately separate three sound sources from different orientations in a semi-anechoic chamber environment, and the extracted target speech has high clarity and intelligibility.