Neural Optimization of Low-Rank MVDR Beamforming Filters for Speech Enhancement

Microphone array observations with a large number of microphones often exhibit low-rank characteristics. Effectively recognizing and leveraging this low-rank structure is essential for advanced microphone array processing, which motivates the development of low-rank beamforming techniques. Such beamformers are based on Kronecker product decomposition, and traditionally, the two sets of sub-filters are estimated through iterative algorithms. In this paper, we propose a neural low-rank beamformer that employs two neural networks to directly estimate the two sets of sub-filters. Each set is predicted by a Conformer-based network, which combines the strengths of Transformers and convolutional neural networks to capture both long-range dependencies and local features. Unlike prior neural Kronecker beamformers, which are limited to standard rectangular array topologies and first-order designs, the proposed method supports arbitrary array topologies and high-order low-rank beamformers. Experimental results demonstrate that the proposed low-rank neural minimum variance distortionless response (MVDR) beamformer consistently outperforms both conventional and Kronecker decomposition-based MVDR beamformers.