A novel approach for underdetermined blind speech sources separation

In this paper, we discussed the separation of n sources from m linear mixtures when the underlying system is underdetermined, that is, when m < n. The underdetermined blind sources separation has two steps. In matrix-recovery step, we defined a characteristic of the speech sources as the durative-sparsity and proposed a novel approach for speech sources separation, which is called as a searching-and-averaging-based method in time domain. This approach tells us how to search some data points which are very close to the basis lines along the direction of basis vectors and how to use them to estimate the mixing matrix. In source-recovery step, we use Bofill and Zibulevsky's shortest-path algorithm. Finally, the experiments results of three speech sources demonstrate the performance of our approach.