ADT Network: A Novel Nonlinear Method for Decoding Speech Envelopes From EEG Signals.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Objective assessment of auditory processing is crucial for diagnosing hearing impairments.
• Current methods for decoding speech envelopes from electroencephalogram (EEG) signals lack both high accuracy and interpretability.
• Developing advanced computational models can enhance our understanding of neural responses to speech.

Purpose of the Study:

• To propose a novel deep learning model, the Auditory Decoding Transformer (ADT) network, for accurate and interpretable speech envelope reconstruction from EEG signals.
• To evaluate the performance of the ADT network against existing nonlinear models.
• To analyze the neural mechanisms underlying speech envelope tracking using the ADT network's interpretability features.

Main Methods:

• Utilized a deep learning approach combining spatio-temporal convolution for feature extraction and a transformer decoder for speech envelope decoding.
• Implemented anticausal masking to ensure the model considers only current and future EEG features, mimicking natural speech-EEG relationships.
• Visualized spatio-temporal convolution weights as time-domain filters and brain topographies, alongside an ablation study of temporal convolution kernels.

Main Results:

• Achieved competitive average reconstruction scores of 0.168 (SparrKULee) and 0.167 (DTU), comparable to other nonlinear models.
• Identified low- (0.5-8 Hz) and high-frequency (14-32 Hz) EEG signals as critical for speech envelope reconstruction.
• Demonstrated that active brain regions are primarily bilateral within the auditory cortex, aligning with prior research.
• Attention score visualizations further corroborated existing findings in auditory processing research.

Conclusions:

• The Auditory Decoding Transformer (ADT) network offers a balanced approach to high-performance and interpretable speech envelope decoding from EEG.
• The model's interpretability provides insights into neural speech envelope tracking mechanisms.
• The ADT network presents a promising tool for objective auditory processing assessment and potential advancements in hearing loss diagnosis.