Predicting speech intelligibility from EEG in a non-linear classification paradigm

Area of Science:

• Neuroscience
• Biomedical Engineering
• Speech Science

Background:

• Current speech intelligibility tests rely on behavioral responses, limiting their use in certain populations.
• Objective measures are needed to assess speech intelligibility, especially for individuals unable to participate actively.
• Brain imaging, specifically electroencephalography (EEG), offers a potential avenue for objective speech intelligibility assessment.

Purpose of the Study:

• To develop and validate a deep-learning model for objective speech intelligibility prediction using EEG.
• To assess the model's performance across various parameters like input segment length and EEG frequency bands.
• To establish a correlation between the model's predictions and a gold-standard behavioral speech intelligibility test.

Main Methods:

• A deep-learning model with dilated convolutions was developed, operating on a match/mismatch paradigm.
• Model performance was evaluated against baseline models, varying input segment length, EEG frequency bands, and receptive field sizes.
• The model's accuracy was assessed on held-out data, with and without fine-tuning, and correlated with the MATRIX test results.

Main Results:

• The dilated convolutional model significantly outperformed baseline models across tested parameters.
• Fine-tuning the model notably improved accuracy on unseen data.
• A significant correlation (r=0.59, p=0.0154) was observed between the model's predicted speech reception threshold (SRT) and the behavioral MATRIX test results.

Conclusions:

• The developed deep-learning model provides an objective measure of speech intelligibility using EEG.
• This method enables prediction of SRT from EEG for subjects without prior training, a novel contribution.
• The findings pave the way for more accessible and objective speech intelligibility assessments.