Speech prediction of a listener via EEG-based classification through subject-independent phase dissimilarity model

Abstract

Abstract This study examines the consistency of cross-subject electroencephalography (EEG) phase tracking in response to auditory stimuli via speech classification. Repeated listening to audio induces consistent EEG phase alignments across trials for listeners. If the phase of EEG aligns more closely with acoustics, cross-subject EEG phase tracking should also exhibit significant similarity. To test this hypothesis, we propose a generalized subject-independent phase dissimilarity model, which eliminates the requirement for training on individuals. Our proposed model assesses the duration and number of cross-subject EEG-phase-alignments, influencing accuracy. EEG responses were recorded from seventeen participants who listened three times to 22 unfamiliar one-minute passages from audiobooks. Our findings demonstrate that the EEG phase is consistent within repeated cross-subject trials. Our model achieved an impressive EEG-based classification accuracy of 74.96%. Furthermore, an average of nine distinct phasic templates from different participants is sufficient to effectively train the model, regardless of the duration of EEG phase alignments. Additionally, the duration of EEG-phase-alignments positively correlates with classification accuracy. These results indicate that predicting a listener’s speech is feasible by training the model with phasic templates from other listeners, owing to the consistent cross-subject EEG phase alignments with speech acoustics.