SpatialNet-Echo: Real-Time Acoustic Echo Cancellation via Integrated Narrow-Band and Cross-Band Processing

Acoustic echo cancellation (AEC) remains challenging in real-time communications due to narrow-band processing limitations and nonlinear distortions. This paper proposes SpatialNet-Echo, the first end-to-end model integrating narrow-band temporal modeling with cross-band spectral coherence for real-time AEC. Our architecture incorporates time-frequency convolution block (TFCB) to capture joint spectrotemporal features via depthwise convolutions, squeeze-and-excitation (SE) blocks to dynamically weight critical channels, and Mamba-based narrow-band processors for efficient long-context modeling. A hybrid loss function jointly optimizes scale-invariant signal-to-noise ratio (SI-SNR), magnitude spectrum, and real-imaginary components to ensure phase coherence. Evaluated on ICASSP 2023 AEC Challenge blind test, SpatialNet-Echo achieves state-of-the-art (SOTA) 4.81 EMOS in far-end single-talk and competitive 4.05 DMOS in double-talk. The lightweight variant further demonstrates deployability.