Thu-2-1-1 Exploiting Conic Affinity Measures to Design Speech Enhancement Systems Operating in Unseen Noise Conditions

Abstract: Speech enhancement under unseen noise conditions is a challenging task, but essential for meeting the increasing de- mand for speech technologies to operate in diverse and dynamic real world environments. A method that has been widely used to enhance speech signals is nonnegative matrix factorization (NMF). In the training phase NMF produces speech and noise dictionaries which are represented as matrices with nonnega- tive entries. The quality of the enhanced signal depends on the reconstruction ability of the dictionaries. A geometric interpre- tation of these nonnegative matrices enables us to cast them as convex polyhedral cones in the positive orthant. In this work, we employ conic affinity measures to design systems able to operate in unseen noise conditions, by selecting an appropriate noise dictionary amongst a pool of potential candidates. We show that such a method yields results similar to those that would be produced if the oracle noise dictionary was used.