Speech Enhancement via Maximum Likelihood Modal Beamforming with Complex Gaussian and Laplacian Priors

Capturing speech from a target source in the presence of interfering sources and noise using acoustic beamforming is an important processing tool for machine listening. When the target speaker can be at any location in the 3D space, spherical microphone arrays are desirable due to their ability to steer the beamformer towards any direction without affecting its directivity pattern. In this work, two beamformers based on the maximum likelihood estimation principle are introduced in the spherical harmonics domain. The first beamformer is designed by assuming that the coefficients of the target speech in the time-frequency domain at the beamformer output follow a zero-mean complex Gaussian prior distribution with time-varying variances. As speech coefficients are better modelled using Laplacian distribution, the second beamformer is designed by assuming a Laplacian prior for the target speech coefficients. To aid the capture of the desired speech, a distortionless constraint is also added to the formulation of both beamformers. The iterative update rules for the variances and the weights of both beamformers have been derived. Simulation results show that the proposed beamformers are more effective in target speech enhancement and distant speech recognition applications. � 2023 Elsevier B.V., All rights reserved.