Frequency domain multi-channel noise reduction based on the spatial subspace decomposition and noise eigenvalue modification

In this paper, frequency domain multi-channel noise reduction algorithms are proposed, based on the subspace decomposition of narrow-band spatial covariance matrices. In speech-present periods, the multi-channel input signals are decomposed into speech and noise spatial subspaces. The noise eigenvalues are modified in order to update the noise statistics not only in the noise-only period but also in the speech-present period. Three approaches are introduced for the noise eigenvalue modification, which are based on the rank-1 property of the speech narrow-band spatial covariance matrix for the single speech source. The proposed algorithms are tested with the simulated data and real data, and the results show that the proposed methods yield better performance compared to the conventional multi-channel Wiener filtering and the time domain subspace approaches.