Late reverberant spectral variance estimation using acoustic channel equalization

In many single- and multi-channel speech dereverberation methods an estimate of the late reverberant spectral variance (LRSV) is required. Contrary to LRSV estimators based on room acoustical properties, such as reverberation time, or based on isotropic models of the reverberant sound field, in this paper we propose to use acoustic channel equalization with estimated room impulse responses (RIRs) for LRSV estimation. Unlike the typical application of acoustic channel equalization, where the objective is to estimate the anechoic or the early reverberant speech component, here the late reverberant part of the estimated RIR is set as the target response. The combination of the proposed LRSV estimator with a beamformer and a spectral gain aims at a tradeoff between the performance of acoustic channel equalization and the robustness of methods based on models of the reverberant sound field. The performance, evaluated for different levels of RIR estimation error, is compared to the results obtained using a maximum likelihood estimator (MLE) of the LRSV, based on an isotropic model of the reverberant sound field, and to a state-of-the-art acoustic channel equalization method. Experimental results for different acoustic scenarios show that for medium levels of RIR estimation errors the proposed method outperforms acoustic channel equalization as well as the maximum-likelihood LRSV estimator in terms of instrumental speech quality measures.