Continuous speech prediction from auditory EEG via exclusive region segmentation

Synthesizing full speech from electroencephalography (EEG) signals is a challenging task. In this study, we investigate the reconstruction of continuous speech that a listener has perceived from non-invasive EEG using a Transformer-based deep learning model. Rather than utilizing the entire brain signal, we examine which brain regions yield more accurate extraction of speech features related to the perceived speech. Previous research had included overlapping electrode placements across regions; in this study, we introduced an exclusive, non-overlapping regional segmentation to enable a more in-depth investigation, thereby allowing clearer identification of each region’s unique information contribution. We configured electrodes as non-overlapping segments and conducted validation experiments to extract speech features from each region that are necessary for speaker characterization and phoneme prediction. [Work supported by JSPS Grant 24H00715 and JST SPRING Grant JPMJSP2131.]