Blind separation of nonstationary sources by block decorrelation of output signal

摘要

A microphone system which separates the objective sound from background noises using the blind source separation technique is under investigation. There are many studies handling the blind source separation problems for nonstationary sources. Among them, Matsuoka, et. al. proposed a neural network approach which decorrelates the outputs of the network with each other over every time. This paper describes a new method which divides the observation signals into blocks and calculates in each block a set of separation matrices so as to decorrelate the output signals with each other. Furthermore, it shows that the intersection of the sets converges to the target set, which can separate the source signals, with increasing number of blocks.