A new regularized transform-domain NLMS adaptive filtering algorithm

The transform domain normalized LMS (TD-NLMS)-adaptive filtering algorithm is an efficient adaptive filter with fast convergence speed and reasonably low arithmetic complexity. However, it is sensitive to the level of the excitation signal, which may vary significantly over time in speech and audio signals. This paper proposes a new regularized transform domain NLMS (R-TDNLMS) algorithm and studies its mean and mean square convergence performance. The proposed algorithm extends the conventional TDNLMS algorithm by imposing a regularization term on the coefficients to reduce the variance of the estimator. The mean and mean square convergence behaviors of the proposed algorithm are studied to characterize its convergence condition and steady-state excess mean squares error (MSE). It shows that regularization can help to reduce the MSE for coloured inputs by trading slight bias for variance. Moreover, the immunity to varying input signal level is significantly reduced. Computer simulations are conducted to examine the effectiveness of the proposed algorithm and they are in good agreement with the theoretical analysis.