Auditory Spatial Attention Detection Based on Ear Eeg Signals Using a Resnet with Channel Attention

Abstract

In recent years, extensive research has focused on enhancing the performance of Auditory Spatial Attention Detection (ASAD) in cocktail party scenarios using electroencephalogram (EEG) signals, given its potential for applications such as AI-driven earphones and neural-steered hearing aids. However, most ASAD methods rely on scalp EEG signals, which pose challenges in portability and discretion, limiting their practicality in real-world settings. To overcome these limitations, this study introduces CAResNet, a novel neural network model designed to decode auditory attention using EEG signals from periauricular electrodes. CAResNet integrates a channel attention module and a residual neural network to enhance ASAD performance. The channel attention module dynamically assigns weights to different EEG channels, amplifying signals from brain regions relevant to ASAD while suppressing irrelevant noise. Meanwhile, the residual neural network automatically extracts deep auditory attention features from the EEG signals. The model’s effectiveness is first evaluated on a full-scalp EEG dataset (KUL dataset), demonstrating state-of-the-art performance with decision windows ranging from 0.1s to 2s. Further analysis on 16-channel periauricular EEG signals from the KUL dataset achieves an impressive accuracy of over 88% within a 0.5s decision window. Lastly, experiments on a dedicated periauricular EEG dataset (cEEGrid dataset) yield accuracy rates exceeding 97% across decision windows from 0.1s to 2s. These findings validate the feasibility of ASAD using periauricular EEG signals and underscore its potential for real-world applications. Additionally, the visualization of the channel attention mechanism reveals brain patterns that align with established neuroscientific findings, further supporting the model’s effectiveness.

Keywords: Ear EEG, Cocktail Party, Auditory Spatial Attention Detection, Channel Attention, Resnet