A PEM-based frequency-domain Kalman filter for adaptive feedback cancellation

Abstract

Adaptive feedback cancellation (AFC) algorithms are used to solve the problem of acoustic feedback, but, frequently, they do not address the fundamental problem of loudspeaker and source signal correlation, leading to an estimation bias if standard adaptive filtering methods are used. Loudspeaker and source signal prefiltering via the prediction-error method (PEM) can address this problem. In addition to this, the use of a frequency-domain Kalman filter (FDKF) is an appealing tool for the estimation of the adaptive feedback canceler, given the advantages it offers over other common techniques, such as Wiener filtering. In this paper, we derive an algorithm employing a PEM-based prewhitening and a frequency-domain Kalman filter (PEM-FDKF) for AFC. We demonstrate its improved performance when compared with standard frequency-domain adaptive filter (FDAF) algorithms, in terms of reduced estimation error, achievable amplification and sound quality.