TFSA-Net: Structured learning for broadband DOA estimation in adverse acoustic conditions

This paper presents TFSA-Net, a data-driven framework for broadband direction-ofarrival (DOA) estimation that is specifically designed to operate under low signal-tonoise ratios (SNRs), limited snapshots, and inter-subband inconsistencies. TFSA-Net employs multi-scale convolutional blocks and gated recurrent units (GRUs) to extract robust temporal-frequency features and implicitly reconstruct spatial coherence across frequency bands. Rather than relying on idealized statistical assumptions or pursuing closed-form estimation bounds, the network learns to perform adaptive frequency focusing and subspace-aware filtering in a physically constrained manner. A regression head is introduced to predict DOAs from the estimated spatial spectrum, preserving compatibility with classical MUSIC-based intuition. Additionally, a permutationinvariant unit-vector loss (PIRMSE-U) is designed to address angle periodicity and multi-source ambiguity. Extensive experiments across simulated broadband acoustic scenarios confirm that TFSA-Net delivers consistent gains in estimation accuracy, robustness, and generalization over traditional subspace methods and recent deeplearning models.