Post-nonlinear speech mixture identification using single-source temporal zones & curve clustering

In this paper, we propose a method for estimating the nonlinearities which hold in post-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. The main novelty of our work consists of using nonlinear single-source confidence measures and curve clustering. Such an approach may be seen as an extension of linear instantaneous sparse component analysis to post-nonlinear mixtures. The performance of the approach is illustrated with some tests showing that the nonlinear functions are estimated accurately, with mean square errors around 4e-5 when the sources are “strongly” mixed.