Developing Lightweight Models with Data Optimization for Attending Speaker Identity from EEG without Spatial...

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Spatial auditory attention decoding (Sp-AAD) is crucial for brain-computer interfaces (BCIs).
• Previous Sp-AAD methods often rely on eye gaze artifacts, not true auditory attention.
• This research investigates EEG signal discriminability for target speaker identity, excluding eye gaze influence.

Purpose of the Study:

• To verify if EEG signals possess sufficient features for decoding auditory attention to a specific speaker.
• To develop and validate a method to mitigate eye gaze artifacts in Sp-AAD.
• To create a computationally efficient model for Sp-AAD.

Main Methods:

• Proposed an EEG-Mixup data optimization technique to adjust data distribution and generate soft labels, suppressing trial-specific features.
• Developed a lightweight EEG-MLP model with approximately 2.5k parameters.
• Evaluated model performance against the state-of-the-art DenseNet-3D model in cross-trial scenarios.

Main Results:

• The EEG-Mixup method significantly improved model generalization without increasing data volume.
• The lightweight EEG-MLP model outperformed DenseNet-3D in cross-trial performance.
• The proposed model demonstrated enhanced computational efficiency and inference speed.

Conclusions:

• EEG signals contain discriminative features for target speaker identity decoding, independent of eye gaze.
• Data optimization techniques like EEG-Mixup can enhance BCI performance.
• Lightweight models offer a practical and efficient approach for future auditory BCI systems.