A comparison of robust capon beamformers using a large-scale microphone array for speech extraction

Abstract

The Capon beamformer has been widely studied in narrowband applications and has already been extended to wideband signals, such as speech. For beamforming-based speech extraction applications, it is still a hot topic in improving the robustness and reducing the computational load, especially when the number of microphones is relatively large. This is because the Capon beamformer with more microphones becomes more sensitive to signal model errors and its computational complexity also increases dramatically. To improve the robustness, a variety of robust Capon beamformers have already been proposed for narrowband signals in recent years, while it lacks a comprehensive evaluation of these beamformers with a large-scale microphone array in noisy and reverberant environments for speech extraction. This paper focuses on evaluating some representative Capon beamformers by simulations and practical experiments. Besides, a low computational complexity robust Capon beamformer is proposed by integrating one-bit quantization and the Kronecker product into a complex generalized Gaussian distribution based maximum likelihood distortionless response beamformer. Simulation and experimental results show that robust Capon beamformers frequently outperform the standard Capon beamformer and traditional fixed beamformers in terms of both objective and subjective measures. It is also demonstrated that the newly proposed beamformer is more computationally efficient and outperforms other existing robust Capon beamformers.