Experimental study of robust acoustic beamforming for speech acquisition in reverberant and noisy environments

The performance of adaptive beamformers may suffer from significant degradation in the presence of steering vector errors, statistics estimation errors, and reverberation. To address this issue, robust beamforming methods, which were originally studied in the narrow-band cases, are studied and compared in this paper for processing acoustic and speech signals, which are broadband in nature. We study two types of methods. In the first type, the robustness of the beamformer is improved by adding a norm constraint and/or a steering vector uncertainty constraint to the optimization problem. It is worth noticing that the norm constraint also helps to control the sidelobes of the beampattern, which makes the beamformers able to suppress interference and multipath effect, thereby improving the robustness of the beamformers with respect to reverberation. Another type of methods improve the robustness using the spatial smoothing technique in which the noise covariance matrix is implicitly estimated first by subtracting an estimate of the desired signal covariance matrix with a delay-and-sum beamformer from the observation signal covariance matrix. Simulations and experiments are performed to investigate the performance of the studied robust adaptive beamformers in acoustic environments. The results show that the robust beamformers outperform their non-robust counterparts in terms of: (1) better performance in reverberation and different noise levels; (2) resilience against steering vector and noisy signal covariance matrix estimation errors; and (3) better predicted speech quality and intelligibility measured using the PESQand STOI measures. (C) 2020 Elsevier Ltd. All rights reserved.