Compressed sensing based underdetermined blind source separation with unsupervised sparse dictionary self-learning

In this paper, we propose an unsupervised blind source separation (BSS) method with sparse dictionary self-learning for the solution of Compressed Sensing (CS) based BSS problem. The idea is to convert BSS problem into general CS form and incorporate an iterative dictionary self-learning strategy with sparse reconstruction to solve it. The proposed method contains two stages. Firstly, a feature clustering method is used to estimate the mixing matrix. Secondly, an alternating iteration procedure is introduced to refine the estimation of the sparse dictionary and the source signals. In each iteration step, the sparse dictionary is obtained using a self-learning algorithm with the last estimates of the sources signals, and then the sources signals are reestimated using the CS based blind separation method with the new sparse dictionary. By adaptively regenerating the dictionaries, the refined dictionaries are approaching the optimal sparse basis of the original sources, which results in the separation performance improvement simultaneously. This dictionary self-learning method doesn´t need any prior information about the original speeches, i.e., it is an unsupervised method. Simulation results show that the proposed method outperforms several state-of-the-art algorithms in either free-noise or noisy case.