AADNet: An End-to-End Deep Learning Model for Auditory Attention Decoding.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Auditory attention decoding (AAD) identifies attended speech in multi-talker environments using electroencephalography (EEG).
• Current AAD algorithms typically use a two-step approach based on neural entrainment to speech envelopes.
• AAD has significant potential for neuro-steered hearing devices.

Purpose of the Study:

• To propose and evaluate AADNet, a novel end-to-end deep learning architecture for auditory attention decoding.
• To compare AADNet's performance against traditional stimulus decoding methods.
• To assess AADNet's effectiveness in both subject-specific and subject-independent scenarios.

Main Methods:

• Developed AADNet, an end-to-end neural network combining prediction and identification stages.
• Compared AADNet with linear stimulus reconstruction, canonical correlation analysis, and non-linear stimulus reconstruction.
• Validated the approach using three distinct electroencephalography (EEG) datasets.

Main Results:

• AADNet demonstrated significant performance improvements over traditional methods.
• Achieved high subject-independent classification accuracies, ranging from 56.3% to 89.4% across datasets and window lengths.
• Showed a superior ability to generalize to data from unseen subjects.

Conclusions:

• The proposed AADNet offers a more direct and effective approach to auditory attention decoding.
• Deep learning models show great promise for advancing AAD technology.
• Results have promising implications for the development of future hearing aids and assistive devices.