A Time-Sequence Decoding Network Based on Frequency Band Power Characteristics for Auditory Attention Decoding

The human ability to locate and track a target sound source in a noisy environment, known as the 'cocktail party effect', is a well-studied phenomenon. Auditory attention decoding (AAD) using electroencephalography (EEG) has emerged as a prominent research focus in this area. Deep learning has significantly advanced AAD performance, yet traditional AAD decoders still leave room for improvement. In this paper, we enhance a widely-used end-to-end ShallowConvNet model (AAD-ShallowConvNet) to improve the detection of auditory spatial attention from EEG signals. ShallowConvNet is traditionally employed to decode EEG band power features. For this study, we optimized the parameters and modules of ShallowConvNet using AAD data. In AAD experiments utilizing a publicly available dataset, AAD-ShallowConvNet demonstrated outstanding performance across different channels and time windows. Under various evaluation conditions, AAD-ShallowConvNet outperformed other state-of-the-art algorithms.