A Robust RLS Implementation of the ANC Block in GSC Structures

Adaptive beamforming, and the minimum variance distortionless response (MVDR) beamformer in particular, is widely used in speech enhancement applications. We consider the enhancement of a single desired speaker in a vehicle (e.g. the driver), in road noise environment. For our problem of a fixed look direction and continuous noise tracking, the generalized sidelobe canceler (GSC) decomposition was shown to be computationally efficient for the implementation of the MVDR criterion. To address robustness issues that arise due to array imperfection, the projected least mean squares (LMS) algorithm, commonly adopted for the adaptive noise canceler (ANC) block realization, was shown to be effective. Here, due to the high dynamics of vehicle road and cabin noises, we propose utilizing a recursive least squares (RLS) flavor algorithm for the realization of the ANC block. To address the robustness issue, we introduce the Modified RLS algorithm. The Modified RLS converges to the required amount of diagonal loading which is associated with the array immunity to imperfection. The proposed diagonal loading algorithm can be easily employed in any diagonal loading problems utilizing RLS based adaptive filtering (AF). We present its operation considering audio signals recorded in a vehicle, demonstrating the way artifacts are mitigated when applied.