Multi-scale dynamic cooperative attention network for auditory attention detection

Humans exhibit a remarkable capacity to concentrate on particular auditory inputs amid multiple simultaneous speakers, as seen in cocktail party settings. Auditory Attention Decoding (AAD) approaches seek to identify which conversation a listener is engaged with by examining neural activity, especially electroencephalography (EEG) recordings. Still, current techniques encounter notable difficulties, although they frequently incorporate supplementary cues like spatial or frequency-domain attributes to boost decoding performance, they commonly overlook correlations among different EEG channels. In this paper, we propose a neuroscience-based architecture for decoding auditory attention, aiming to achieve structured spatial interaction modeling across electrode groups and effective multi-scale feature integration. The architecture comprises two complementary modules: (1) a Dynamic Feature Fusion Attention Module (DFFAM) that employs attention-weighted multi-scale convolutions to adaptively capture subject-specific temporal features without manual kernel size tuning; (2) a dual cross-self attention module that explicitly models both intra-group and inter-group electrode interactions through a structured four-branch attention mechanism, motivated by the hemispheric lateralization of auditory spatial attention. Comprehensive experiments on two public datasets demonstrate that the proposed method achieves state-of-the-art performance, with particular advantages in cross-subject robustness and ultra-short decision window scenarios. Code is available at:https://github.com/jiezishu737/code-for-eeg-singal.