Dynamic Attention-Asymmetric Perceptron Network for Overlapping Sound Event Detection

Detecting overlapping sound events in complex acoustic environments presents a significant challenge due to the intricate variations in both temporal and spectral domains, which hinders the distinction of individual events, especially under low SNR conditions. To address this challenge, we propose the dynamic attention asymmetric perceptron network (DAAPNet), which integrates three novel modules: bidirectional asymmetric convolution (BA-Conv), time frequency dynamic perceptron (TFDP), and dynamic dense synthesizer attention (DDSA). The BA-Conv module uses asymmetric convolution kernels to capture temporal and spectral patterns in separate pathways, helping to separate overlapping acoustic features while maintaining computational efficiency. The TFDP module processes acoustic signals with three stages: expansion, grouping, and cross-shifting perception, to adapt to variations in both the time and frequency domains. The DDSA mechanism is developed by using multi-scale context sizes with learnable fusion weights, making it effective for modeling both brief and extended acoustic events. Evaluated on whale recordings and the public DCASE dataset, each under multiple SNR levels, DAAPNet outperforms state-of-the-art methods by 10.7$\%$in PSDS at very low SNR while achieving improved performance with lower computational cost on more complex data. The source code of DAAPNet is available athttps://github.com/OceanLab-AHU/North_whale_dcase_open.