The NERCSLIP-USTC System for Track2 of the First Chinese Auditory Attention Decoding Challenge

The human brain has the ability to selectively focus on a specific auditory stimulus in multi-speaker scenarios. Recently, it was shown that clues on the target speaker can be extracted from electroencephalography (EEG) signals, which is termed as auditory attention decoding (AAD). Existing AAD methods mainly use EEG data elicited by audio-only stimuli and are evaluated on the subjects that have appeared in the training set, resulting in a poor generalization capacity. In the ISCSLP 2024 Chinese AAD Challenge, we explore the performance of AAD models on audio-only and audio-video scenarios. We evaluate 3 different methods and 3 dataset splitting and training strategies and propose to utilize the DenseNet-3D approach and cross-subject data utilization as the submitted system. The effectiveness is verified in the cross-subject track of this challenge, and our method ranks the first with an average accuracy of 59.91%. The reproducible source code is available at: https://github.com/zyb6239/my-AAD-challenge.