Embedded optimization algorithms for multi-microphone dereverberation

In this paper we propose a new approach to multi-microphone dereverberation, based on the recent paradigm of embedded optimization. The rationale of embedded optimization in performing online signal processing tasks, is to replace traditional adaptive filtering algorithms based on closed-form estimators by fast numerical algorithms solving constrained and potentially non-convex optimization problems. In the context of dereverberation, we adopt the embedded optimization paradigm to arrive at a joint estimation of the source signal of interest and the unknown room acoustics. It is shown how the inherently non-convex joint estimation problem can be smoothed by including regularization terms based on a statistical late reverberation model and a sparsity prior for the source signal spectrum. A performance evaluation for an example multi-microphone dereverberation scenario shows promising results, thus motivating future research in this direction.