Data-driven machine learning models for decoding speech categorization from evoked brain responses

Speech categorization is better described by induced rather than evoked neural activity

Categorical perception (CP) describes how the human brain categorizes speech despite inherent acoustic variability. We examined neural correlates of CP in both evoked and induced electroencephalogram (EEG) activity to evaluate which mode best describes the process of speech categorization. Listeners labeled sounds from a vowel gradient while we recorded their EEGs. Using a source reconstructed EEG, we used band-specific evoked and induced neural activity to build parameter optimized support vector machine models to assess how well listeners' speech categorization could be decoded via whole-brain and hemisphere-specific responses. We found whole-brain evoked β-band activity decoded prototypical from ambiguous speech sounds with ∼70% accuracy. However, induced γ-band oscillations showed better decoding of speech categories with ∼95% accuracy compared to evoked β-band activity (∼70% accuracy). Induced high frequency (γ-band) oscillations dominated CP decoding in the left hemisphere, whereas lower frequencies (θ-band) dominated the decoding in the right hemisphere. Moreover, feature selection identified 14 brain regions carrying induced activity and 22 regions of evoked activity that were most salient in describing category-level speech representations. Among the areas and neural regimes explored, induced γ-band modulations were most strongly associated with listeners' behavioral CP. The data suggest that the category-level organization of speech is dominated by relatively high frequency induced brain rhythms.