Télécom Paris

MSE optimal regularization of APA and NLMS algorithms in room acoustic applications
  • Toon van Waterschoot, (Katholieke Universiteit Leuven, ESAT-SCD)
  • Geert Rombouts, (Katholieke Universiteit Leuven, ESAT-SCD)
  • Marc Moonen, (Katholieke Universiteit Leuven, ESAT-SCD)
  • Adaptive filtering algorithms and structures for echo and noise control
Get the paper in PDF format
Acrobat Reader (version 5 minimum) is necessary to read this document.

A common way of increasing the robustness of affine projection and normalized least mean squares adaptive filtering algorithms, is to add a scaled identity regularization matrix to the input signal correlation matrix before inversion. This ad-hoc method can also be interpreted as the result of minimizing a regularized underdetermined least squares criterion. Moreover, by relating this criterion to linear minimum mean square error estimation, we can derive MSE optimal APA and NLMS algorithms, which feature a regularization matrix that is not necessarily a scaled identity matrix. The proposed algorithms allow for incorporating prior knowledge on both the near-end signal and the true room impulse response, and are intimately linked to Levenberg-Marquardt regularization and proportionate adaptation. Simulation results of echo and feedback cancellation experiments confirm that the adaptive filter convergence speed and tracking properties may be considerably improved using the proposed algorithms.

©2006 Télécom Paris/TSI
Edition : Télécom Paris -- 2006