A robust NLMS algorithm with a novel noise modeling based on stationary/nonstationary noise decomposition

This paper proposes a robust NLMS algorithm with a novel noise modeling based on stationary/nonstationary noise decomposition. The ambient noise including the near-end signal is modeled as a weighted sum of the stationary and the nonstationary components. These components are independently estimated with an appropriate time constant for better accuracy. The estimates are weighted by the stationary/ nonstationary likelihood before summation. The integrated noise estimate controls the coefficient adaptation stepsize such that it is an upward convex function of the reference input with a noise offset for robustness. Evaluations in a fullband and a subband echo cancellation scenarios show that ERLE has been improved by as much as 40 dB over the algorithm with a conventional noise model in both single- and double-talk sections with no double-talk detection.