关键词: 听觉注意力检测, 脑电信号, 图神经网络, 时空频融合机制, 解码精度

Abstract: Normal humans can focus on specific speakers in multi-speaker environments, and Auditory Attention Detection (AAD) aims to analyze EEG decoding characteristics related to the attended speaker's speech waveform to model auditory attention selection. However, existing AAD methods are limited by single-domain analyses, neglecting the interplay between time, frequency, and spatial information. Graph Neural Network (GNN) performs well in processing spatial non-Euclidean data. Based on this, a dual-branch network architecture of spatial-temporal frequency fusion for high-speed and efficient auditory attention detection is studied in this paper. The network includes spatial-temporal attention branch and frequency attention branch. The former uses Transformer to capture global context information and GNN to model local spatial topology; the latter uses residual convolutional network to extract multi-band EEG spectral features. Finally, the two branches are fused to comprehensively consider time, space and frequency features. The classification results of auditory attention detection were obtained. Validation on the KUL dataset showed significant improvements in decoding accuracy, achieving 88.75% and 95.31% within 0.1s and 1s decision windows, respectively, and 94.88% in the 5s window, confirming the efficacy of our proposed time-space-frequency attention mechanism.

Key words: auditory attention, electroencephalogram, graph neural network, time-spatial-frequency attention, decoding accuracy