Convex combinations of adaptive filters for feedback cancellation in hearing aids

Abstract

In real-time applications, such as acoustic feedback cancellation in hearing aids, it is desired that the adaptive algorithm converges fast towards a good estimate of the feedback path, while incurring a low steady-state error. A high adaptation step of the least mean square (LMS) algorithm allows for a faster convergence but compromises the steady-state misalignment of the adaptive filter. In this paper, we propose the use of an affine combination of two adaptive filters with different step sizes, but a common combination parameter to break the precision-versus-speed compromise of the adaptive algorithm. Moreover, an improved version of this affine-combination scheme using a different combination parameter for each filter coefficient is also applied for better tracking. Further, a more sophisticated algorithm, which utilizes a convex combination of adaptive filters in a three-filter scheme along with the varying step size (VSS) approach, is introduced for an improved tracking and faster convergence to eliminate the effect of the feedback path. To reduce the estimation bias due to closed-loop signal correlations, the linear prediction-based adaptive feedback cancellation (AFC) design is used instead of a basic adaptive feedback canceller. Simulation results show that the convex combination schemes provide better feedback-cancellation performance than the single-filter VSS algorithm.