White Rose University Consortium logo
University of Leeds logo University of Sheffield logo York University logo

Acoustic echo cancellation using frequency-domain spline identification

Zakharov, Y.V., Tozer, T.C. and Pearce, D.A.J. (2007) Acoustic echo cancellation using frequency-domain spline identification. IEEE Transactions on Signal Processing, 55 (2). 585 - 593. ISSN 1053-587X

Full text not available from this repository.


Low-complexity delayless acoustic echo cancellation techniques based on frequency-domain spline-identification are proposed and investigated. Two methods of approximation of the acoustic frequency response, both using B-splines, are considered: the optimal-spline method and the local-spline method. The optimal-spline method seeks the solution of a least squares problem. The most computationally demanding part of the method, solution of the normal equations, is implemented by using the low-complexity dichotomous coordinate descent algorithm. The local-spline method avoids solving the normal equations, enabling further simplification; this is at the expense of a slight degradation in the cancellation performance. A novel efficient double-talk detector is also proposed, being an inherent feature of the frequency-domain identification. Open-loop and closed-loop identification schemes with cubic splines are studied by simulation and compared with the fast affine projection (FAP) algorithm. The proposed techniques provide cancellation performance better than that of the FAP algorithm, especially in double-talk and noisy environments, with a lower complexity.

Item Type: Article
Institution: The University of York
Academic Units: The University of York > Electronics (York)
Depositing User: York RAE Import
Date Deposited: 22 May 2009 10:29
Last Modified: 22 May 2009 10:29
Published Version: http://dx.doi.org/10.1109/TSP.2006.887150
Status: Published
Publisher: IEEE
Identification Number: 10.1109/TSP.2006.887150
URI: http://eprints.whiterose.ac.uk/id/eprint/6305

Actions (repository staff only: login required)