Adaptive Feedback Cancellation Using a Partitioned-Block Frequency-Domain Kalman Filter Approach With PEM-Based Signal Prewhitening

Subjective and Objective Sound-Quality Evaluation of Adaptive Feedback Cancellation Algorithms

“…First, we present the results of a perceptual subjective sound-quality evaluation of three different AFC algorithms in order to validate the hypothesis of superior performance of the PEM-FDKF over the PEM-FDAF and the block normalized least mean squares (BNLMS) algorithm, as observed in simulations [14]. Second, the test signals used to subjectively evaluate sound quality, are tested using eight different existing objective measures, in order to evaluate the predictive capabilities of these objective measures.…”

“…The resulting algorithm, the PEM-based adaptive feedback canceller (PEM-AFC) [10], yields good convergence performance and limited perceptual distortions, when compared to algorithms employing other decorrelation approaches [1], [11]. Various modifications of the PEM-AFC have been developed by changing some of the underlying assumptions and/or implementation strategies; e. g., among others, the PEM-based adaptive filtering with row operations (PEM-AFROW) [12] makes the algorithm suitable for long feedback paths, the PEMbased frequency-domain adaptive filter (PEM-FDAF) [11], [13] offers a frequency domain implementation of the feedback path adaptation, and the PEM-based frequency-domain Kalman filter (PEM-FDKF) [14] replaces the standard normalized least mean squares (NLMS) type adaptive filter with a Kalman filter.…”

“…Finally, the PEM-FDKF [14] integrates the PEM into the frequency-domain Kalman filter (FDKF) [28], and has been shown to provide superior performance over the PEM-FDAF [14]. The PEM-FDKF exploits the properties of the FDKF, leading to a faster convergence and to an optimal estimation of the step-size parameter [28], simplifying the design choices one has to make in order to use the algorithm.…”

“…Section III-A), or to µ BNLMS = 2 × 10 . The adaptive stepsize of the PEM-FDAF was defined as in [14], with the fixed stepsize µ FDAF = 2 × 10…”

“…The initial value of the error covariance matrix and the transition factor of the PEM-FDKF were set, respectively, to P(0) = 1 × 10 −4 I M ×M and α = 0.999. For more details about the PEM-FDAF and the PEM-FDKF we refer to [25] and [14]. These parameters were chosen such that each algorithm would remain stable for most part of all the source signals employed in the study.…”