Spectral bandpower outperforms statistical features in subject-independent EEG speech decoding

A subject-independent evaluation of imagined speech decoding shows that frequency-domain spectral bandpower pipelines achieve 49.03% mean accuracy, significantly outperforming time-domain statistical features at 37.97%. The study uses a random forest classifier with trial-level majority voting to establish a baseline for cross-subject generalization.