Cross-Layer Hybrid Attention Network for Sound Event Detection and Localization

Abstract

The overlap of polyphonic audio events in real-world sound event detection and localization (SEDL) limits the feature extraction capability of existing neural networks and results in suboptimal performance. To address this limitation, we propose a cross-layer hybrid attention network (CLHANet), a task-oriented cross-layer feature-flow architecture for SEDL. Instead of performing feature fusion only once, CLHANet progressively integrates mel-energy and sound-intensity representations across multiple network depths through cascaded cross-layer feature fusion units. Within each unit, multi-scale local channel information and global spatial context are jointly modeled, while cross-layer connections propagate previously enhanced representations to form a continuous feature path. Hybrid attention, SoftPool, and Conformer are subsequently employed for post-fusion spatial–channel recalibration, information-preserving downsampling, and local–global temporal modeling, respectively. Experiments are conducted on the open dataset of DCASE2022 Task 3. Compared with the baseline CRNN, CLHANet achieves an absolute reduction of 0.05 in <inline-formula> <tex-math notation="LaTeX">$ER_{20^{\circ } }$ </tex-math></inline-formula>, an absolute increase of 13.4 percentage points in <inline-formula> <tex-math notation="LaTeX">$F_{20^{\circ } }$ </tex-math></inline-formula>, an absolute reduction of 6.7° in <inline-formula> <tex-math notation="LaTeX">$LE_{CD}$ </tex-math></inline-formula>, and an absolute increase of 17 percentage points in <inline-formula> <tex-math notation="LaTeX">$LR_{CD}$ </tex-math></inline-formula>.