Parameterized model based blind intrinsic chirp source separation

The blind source separation (BSS) concerns recovering sources from their mixtures. Specifically, the sources in this paper, named intrinsic chirp sources (ICSs), are modeled as the linear combination of non-linear chirp components (NCCs). A novel method is developed here to address the blind separation issue of them. Firstly, all the mixtures at each channel are decomposed into a series of NCCs by a parameterized decomposition approach. It can adapt to NCCs with time-frequency (T-F) distribution suffering from bad T-F concentration and non-disjoint T-F overlapping. Next, the reconstructed NCCs can be clustered into corresponding ICS according to the fact that the NCCs belonging to the identical ICS share the same column in mixing matrix. The source recovery and mixing matrix estimation are finally accomplished based on the clustering consequence. Three simulations demonstrate the capability of our method in dealing with challenging under-determined BSS cases and its potential in practical applications.