Maximum likelihood based noise covariance matrix estimation for multi-microphone speech enhancement

Multi-microphone speech enhancement systems can often be decomposed into a concatenation of a beamformer, which provides spatial filtering of the noisy signal, and a singlechannel (SC) noise reduction filter, which reduces the noise remaining in the beamformer output. Here, we propose a maximum likelihood based method for estimating the intermicrophone covariance matrix of the noise impinging on the microphone array. The method allows prediction of this co-variance matrix for non-stationary noise sources even in signal regions where the target speech signal is present. Although the noise covariance matrix may have several purposes, we use it in this paper for estimating the power spectral density (psd) of the noise entering the SC filter, as this is important for optimal operation of the SC filter. In simulation experiments with a binaural hearing aid setup in a realistic acoustical scenario, the proposed method performs better than existing methods for estimating this noise psd.