Convolutive blind source separation based on multiple times and decorrelation

Mixed signals in practice can be viewed as sums of differently convolved sources, and the signals are non-stationary. The task of blind source separation is to obtain a set of separation filters and make the estimated signals of sources statistically independent. This paper discusses a convolutional blind source separation algorithm based on second-order decorrelation, taking into account non-stationarity of signals. Influence of noise on the quality of separation is considered as well. To avoid inconsistency of frequency bin permutation, a multi-resolution approach to blind source separation is studied. The algorithm is used to separate real acoustic signals successfully. Experimental results are presented and separation performance analyzed. Validity of the algorithm is shown by the improvement of SNR. The algorithm converges rapidly and has high precision. It can be used to separate actual signals recorded in shallow sea.