Authors: Alexis MacIntyre¹; Clément Gaultier²; Tobias Goehring³,¹
¹University of Cambridge; ²Institut Pasteur; ³University of Zurich
Background: During speech perception, acoustic stimulus properties can be reconstructed from EEG signals. While most studies focus on the amplitude envelope, speech acoustics can be characterized by various spectral descriptors. This study assesses how robustly an extended acoustic feature set can be decoded from EEG under different levels of intelligibility and acoustic clarity.
Method: EEG was recorded from 38 young adults listening to intelligible and non-intelligible speech, both unprocessed and spectrally degraded via vocoding. Extracted features included envelope-based descriptors and spectral properties such as instantaneous spectral slope and spectral flux. Decoding robustness was established using linear and nonlinear model architectures, with accuracy standardized using randomly permuted surrogate data.
Results: Linear and nonlinear models yielded similar performance patterns across features and conditions. Z-score conversion revealed noise floor differences between features. While decoding accuracy for some features varied with spectral degradation and intelligibility, these differences were minimized in more robustly decoded features, suggesting reconstruction is primarily driven by generalized auditory processing. Notably, spectral flux uniquely predicted individual comprehension after controlling for acoustic variables.
Conclusion: Linear decoders are as effective as nonlinear models in capturing EEG responses to speech acoustics beyond the amplitude envelope. The reconstruction accuracy of specific features reflects both understanding and spectral clarity, providing insight into how sound properties are represented in the brain and offering potential for future clinical applications.
