Nonlinear blind mixture identification using local source sparsity and functional data clustering

In this paper we propose several methods, using the same structure but with different criteria, for estimating the nonlinearities in nonlinear source separation. In particular and contrary to the state-of-art methods, our proposed approach uses a weak joint-sparsity sources assumption: we look for tiny temporal zones where only one source is active. This method is well suited to non-stationary signals such as speech. We extend our previous work to a more general class of nonlinear mixtures, proposing several nonlinear single-source confidence measures and several functional clustering techniques. Such approaches may be seen as extensions of linear instantaneous sparse component analysis to nonlinear mixtures. Experiments demonstrate the effectiveness and relevancy of this approach.