Cue-Preserving MMSE Filter with Bayesian SNR Marginalization for Binaural Speech Enhancement

Abstract

Binaural speech enhancement has often suffered from the trade-off between noise reduction and spatial cue preservation. The common-gain filtering of noisy speech under minimum mean-square error (MMSE) turned out as a viable approach, which resembles the format of Wiener-filtering spectral enhancement. Those techniques critically require the estimation of the local time-varying a-priori SNR. In single-channel approaches, it has been recently shown that local a-priori SNR can be marginalized in a Bayesian sense with an MMSE approach. In this paper, we translate the single-channel approach into a binaural Bayesian SNR marginalization, based on a binaural a-priori SNR definition and a related hyperprior. The overall MMSE solution then turns into a posterior expectation of an informed cue-preserving Wiener filter function, the computation of which is governed by binaural a-posteriori SNR and global SNR (i.e., the hyper-prior mean). The resulting MMSE solution is thus easy to implement and performance consistently stands at the top of our evaluation by segmental SNR, PESQ, and STOI computational metrics.