12:15pm
EQuad E415
Title: Parametric demixing using second order statistics
Abstract: It has been known for some time that linear mixtures from general filter models of mixing cannot be demixed using only second order statistics. In many applications however, a parametric mixing model, the time delay mixing model, is more appropriate and turns out to be a more accurate description of the physical mixing process. We show that when mixtures are described by a parametric time-delay mixing model second order statistics is sufficient for demixing, provided the sources are statistically independent and the number of emitters is less or equal to the number of receivers. The main reason for this is that statistical independence provides an infinite number of constraints on the finite number of mixing model parameters. The approach does not work for the general filter mixing.
The theoretical results are illustrated in an example of speech demixing in a demo during the presentation.