Enhancing Cross-Subject Generalization for Selective Auditory Attention Decoding Via Multi-Expert Voting

作者：Dahan Wang, Tong Lei, Qinwen Hu, Taihui Wang, Jinzheng Zhao, Rilin Chen, Meng Yu, Dong Yu · 年份：2026 · DOI：10.1109/icassp55912.2026.11461716 · 研究领域：Speech Recognition and Synthesis、Music and Audio Processing、Speech and Audio Processing

Selective Auditory Attention Decoding (AAD) aims to identify the direction of the attended speaker from electroencephalogram (EEG) signals in multi-speaker scenarios. However, existing AAD methods often struggle to generalize to unseen subjects. For the ICASSP 2026 EEG-AAD Challenge, a multi-expert voting strategy, which integrates several state-of-the-art (SOTA) neural networks (NNs) and aggregates predictions from diverse architectures, effectively reducing single-model bias and improving performance and robustness, is proposed. Additionally, direct current (DC) drift removal, rereferencing, and bandpass filter are applied to preprocess the EEG signals. The proposed strategies rank first in track 1 of the EEG-AAD Challenge, achieving an average decoding accuracy of 56.5% on unseen subjects.