Low-Latency Audio Front-End Region-of-Interest Beamforming for Smart Glasses

In wearable audio capture, region-of-interest (ROI) beamforming improves robustness to direction-of-arrival uncertainty while aligning acquisition with the wearer’s field of view. Least-distortion maximum-gain (LDMG) ROI beamformers have been proposed in both the time domain and the short-time Fourier transform (STFT) domain, but their trade-offs for smart-glasses front ends remain unclear. We develop a unified formulation that subsumes both implementations, makes their approximations explicit, and enables a fair, streaming-aware comparison. Using real multichannel recordings from a smart-glasses platform, we quantify algorithmic latency and real-time complexity and benchmark speech enhancement and spatial selectivity. Across all tested conditions, the time-domain implementation delivers higher performance with 2× lower algorithmic latency, at the cost of increased computation. These results indicate that, when low latency is critical and modest additional on-device computing power is available, time-domain ROI beamforming is the preferred choice for smart-glasses front ends.