Adaptive Sparse Source Separation with Application to Speech Signals

In this paper, a sparse component analysis algorithm is presented for the case in which the number of sources is less than or equal to the number of sensors, but the channel (mixing matrix) is time-varying. The method is based on a smoothed ¿⁰ norm for the sparsity criteria, and takes advantage of the idea that sparsity of the sources is decreased when they are mixed. The method is able to separate synthetic and speech data, which require very weak sparsity restrictions. It can separate up to 50 mixed signals while being adaptive to channel variation and robust against noise.