EdgeSepLoc: A Lightweight Multi-Task Network for Real-Time Speech Enhancement and Spectral Localization on Resource-Constrained Devices

Three major contributions: (1) a unified low-latency archi-
tecture for joint enhancement and spectral localization given
a hard constraint on the amount of compute on the edge;
(2) a localization-guided attention gate, leveraging spatial
information to better enhance images, while also improving
the quality of the enhancement; and (3) an uncertainty-aware
confidence head, which flags spatially uncertain bins to avoid
degradation in reverberant environments