Blind Estimation of Acoustic Transfer Functions using Convolutive Transfer Functions with Application to Dereverberation

Array signal processing algorithms based on Acoustic Transfer Functions (ATFs) generally provide better results than Relative Transfer Functions (RTFs). However, direct estimation of ATFs is challenging due to the lack of the source input signal. To tackle this problem, we propose a novel approach for Blind Estimation of ATF based on convolutive transfer functions (CTFs), termed BEAT-C. The method begins by locating the source using the time difference of arrival (TDOA) estimated using the Generalized Cross CorrelationPhase Transform (GCC-PHAT) with a distributed microphone array. Next, the Weighted Prediction Error (WPE) algorithm is used for dereverberation and beamforming, where a uniform linear array is added to the distributed array as a hybrid configuration. With the source signal initially estimated using the Delay and Sum (DAS) beamformer, the CTF coefficients can be computed using either the Wiener filter or the Kalman filter. The simulation results have shown that the proposed method significantly outperforms the adaptive Multichannel Time Domain Least Mean Square (MCLMS) method in the ATF estimation. To further validate the proposed approach, we used the ATF estimates to reduce the room dereverberation using the Multiple Input/Output Inverse Theorem (MINT). The results have shown the superior dereverberation performance achieved by our proposed method over WPE.