Robust decoding of the speech envelope from EEG recordings through deep neural networks

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Auditory attention decoding (AAD) using electroencephalography (EEG) is crucial for enhancing speech comprehension in noisy environments.
• Deep neural networks (DNNs) show potential for improving AAD algorithms, but their real-world robustness is not well understood.

Purpose of the Study:

• To evaluate DNN performance in reconstructing attended speech envelopes from EEG across diverse listening conditions.
• To assess the generalizability of subject-independent DNNs for AAD with unseen participants.

Main Methods:

• Linear models and DNNs were used to decode speech envelopes from EEG data, with and without subject-specific training.
• Models were trained and tested in clean, noisy, and competing-speaker scenarios.
• Performance was evaluated based on reconstruction accuracy and generalizability.

Main Results:

• DNNs significantly outperformed linear models in reconstructing speech envelopes, even without subject-specific data.
• DNNs demonstrated robust performance across various listening conditions and generalized to new datasets.
• Subject-independent DNNs successfully decoded auditory attention in competing-speaker scenarios.

Conclusions:

• DNNs offer a substantial advancement over linear models for AAD, improving speech envelope reconstruction accuracy.
• The generalizability and robustness of DNNs across different conditions make them highly promising for real-world AAD applications.
• DNNs represent a significant step towards developing intelligent hearing aids that can effectively decode user attention.