Binaural LCMV Beamforming with Partial Noise Estimation

Besides suppressing all undesired signal components, another important task of binaural noise reduction algorithms is the preservation of the spatial impression of the acoustical scene, which can be achieved by preserving the binaural cues of all signal components. While the well-known binaural minimum variance distortionless response (BMVDR) beamformer at least can preserve the binaural cues of a single desired source, several extensions have been proposed to additionally preserve the binaural cues of interfering sources and diffuse noise. The binaural linearly-constrained minimum variance (BLCMV) beamformer uses additional constraints to preserve the binaural cues of interfering sources and enables a direct scaling using an interference scaling parameter. The BMVDR with partial noise estimation (BMVDR-N) is aimed at preserving a scaled version of the diffuse noise to partially preserve its interaural coherence and hence its perceived diffuseness. In this paper, we propose to combine both extensions of the BMVDR, leading to the BLCMV with partial noise estimation (BLCMV-N). It is shown that the BLCMV-N can be seen as a mixture of the noisy input signal and the output of a BLCMV that uses an adjusted interference scaling parameter. A theoretical analysis and comparison between the BMVDR, its extensions and the proposed BLCMV-N in terms of noise reduction and binaural cue preservation performance is provided. Experimental results and results of a subjective listening test show that the BLCMV-N is able to to preserve the spatial impression of an interfering source, i.e. like the BLCMV, and yields a trade-off between noise reduction and binaural cue preservation of diffuse noise, i.e. like the BMVDR-N.