A new algorithm for noise PSD matrix estimation in multi-microphone speech enhancement based on recursive smoothing

In this paper, we present a new algorithm for the estimation of the noise power spectral density (PSD) matrix, as needed for multi-microphone speech enhancement in a general non-stationary noisy environment. First, we propose a recursive scheme for noise PSD estimation in which the current, previous and close subsequent noisy speech frames are properly weighted. The forgetting factor for the recursive updating of the smoothed PSD is obtained based on an overall measure of the SNR across all microphone signals. Since this SNR measure depends on the noise statistics, we choose to iteratively update it using the latest available estimate of the noise PSD matrix. Finally, to obtain better estimation accuracy in the proposed method, we further apply a direct extension of the minimum tracking approach to the estimated noise PSD matrix. Performance of the proposed algorithm is evaluated in terms of objective measures and its superiority is shown with respect to two recent noise PSD estimation methods in the context of speech enhancement.