Frequency-invariant beamformer design via ADPM approach

Frequency-invariant (FI) beamformers play an important role in suppressing speech waveform distortions. Due to the perfect FI beampattern (FIB), differential microphone arrays (DMAs) have been widely used in practical applications like voice communication and human-machine interface systems. Superdirective beamforming generated by DMAs have many useful properties but suffer white noise amplification. To address this drawback, we formulate a least squares broadband FI problem, under the white noise out-put power constraints, to improve robustness of superdirective beamformers. The problem is challenging to solve since FI beamformers are designed on broadband and initialization parameters corresponding to each frequency are different. We devise broadband beampattern synthesis algorithm based on alternating direction penalty method (ADPM), which utilizes the relationship between residuals and penalty terms to reduce the iteration number under the improved framework of alternating direction method of multi-pliers (ADMM). The proposed ADPM method can decompose the optimization problem into multi-block convex optimization problems and solve them separately. The fast convergence property of our solution is demonstrated via numerical simulations.(c) 2022 Elsevier B.V. All rights reserved.

Automated summary

Frequency-invariant (FI) beamformers are important in suppressing speech waveform distortions. Differential microphone arrays (DMAs) with perfect FI beampattern (FIB) are widely used in voice communication and human-machine interface systems. However, superdirective beamforming generated by DMAs amplifies white noise. To overcome this drawback, we propose a least squares broadband FI problem under white noise output power constraints, improving the robustness of superdirective beamformers. We develop a broadband beampattern synthesis algorithm based on the alternating direction penalty method (ADPM) to solve the challenging optimization problem, demonstrating fast convergence through numerical simulations. (c) 2022 Elsevier B.V. All rights reserved.