Robust non-vad implementation of Multichannel Wiener filter for binaural noise reduction

Binaural noise-reduction techniques based on Multichannel Wiener filter (MWF) are promising candidates to be used in binaural hearing aids. The real-time implementation of the existing MWF methods involves different challenges, e.g. the robustness against errors in the estimation of the second-order statistics. A typical approach in MWF to estimate these statistics has been the use of a voice activity detector (VAD). But a VAD-based estimation is not particularly suitable for low SNR scenarios, or to track the noise statistics in non-stationary environments during voiced segments. This paper proposes a non-VAD statistics estimation and a method to adapt the trade-off parameter in a well-known MWF method, SDW-MWF. The performance of the proposed implementation is analyzed under realistic scenarios. The proposed method outperforms a VAD-based implementation, is robust under low-SNR and non-stationary environments, and provides good speech quality.