Blind Non-stationnary Sources Separation by Sparsity in a Linear Instantaneous Mixture

摘要

In the case of a determined linear instantaneous mixture, a method to estimate non-stationnary sources with non activity periods is proposed. The method is based on the assumption that speech signals are inactive in some unknown temporal periods. Such silence periods allow to estimate the rows of the demixing matrix by a new algorithm called Direction Estimation of Separating Matrix (DESM). The periods of sources inactivity are estimated by a generalised eigen decomposition of covariance matrices of the mixtures, and the separating matrix is then estimated by a kernel principal component analysis. Experiments are provided with determined mixtures, and shown to be efficient.