S-PHiNe: Physics-Informed Multichannel Speech Enhancement Using Spectro-Spatial Fusion for Low-SNR Conditions

Abstract

Multichannel speech enhancement (MCSE) under adverse noise conditions remains a major challenge. This study proposes S-PHiNe, a neural physics-informed framework integrating deep graph convolutional network (DGCN)-based spatial embedding for steering vector estimation, a complex beamforming neural network (CBNN) for spectral mask estimation, and physics-informed MVDR beamforming within end-to-end physics-constrained training. The model outperforms state-of-the-art (SOTA) neural postfilter and beamforming methods. Evaluations demonstrate how masking, spatial embeddings, and physics-informed constraints jointly improve speech intelligibility, perceptual quality, and multichannel generalization, establishing S-PHiNe as a robust solution for highly noisy multichannel environments.