Decoding of speech acoustics from EEG: going beyond the amplitude envelope

Area of Science:

• Neuroscience
• Auditory Perception
• Signal Processing

Background:

• Speech perception relies on processing acoustic cues.
• Electroencephalography (EEG) can decode brain activity related to speech.
• Current decoding often focuses on the amplitude envelope, neglecting other spectral features.

Purpose of the Study:

• To assess the robustness of decoding an extended set of acoustic speech features from EEG.
• To investigate how intelligibility and spectral degradation affect decoding accuracy.
• To compare linear and non-linear decoding models for speech acoustics.

Main Methods:

• EEG data collected from 38 adults listening to intelligible/non-intelligible, processed/degraded speech.
• Extracted acoustic features including spectral slope and spectral flux, alongside the amplitude envelope.
• Employed multiple linear and non-linear model architectures for decoding, standardizing accuracy with surrogate data.

Main Results:

• Linear models showed comparable or superior performance to non-linear models.
• Decoding accuracy varied across features and conditions, with some features more robust to degradation and intelligibility changes.
• Differences in noise floor between features were observed after scaling accuracy.
• Robustly decoded features indicated generalized auditory processing.

Conclusions:

• Linear decoders are effective for capturing EEG responses to speech acoustics beyond the amplitude envelope.
• Reconstruction accuracy of certain spectral features correlates with speech intelligibility and clarity.
• Findings illuminate differential neural representations of sound properties and suggest potential clinical applications.