Covert speech decoding from EEG cortical current source

 BCIs have been drawn attention to establishing a non-verbal communication system. The most robust method uses evoked EEG signals from mainly visual stimulations using flashing alphabetical letters or computer icons. It would be ideal if BCI can decode any words or letters that users are thinking.

 I have been working on decoding imagined (covert) speech and have shown the possibility of EEG cortical current sources to decode vowel information with higher accuracies than EEG sensor signals (Fig. 1, Frontiers in Neuroscience, 2016). Although the study showed decoding of only two vowels, this was the first success of imagined vowel decoding using EEG data. Further decoders analyses showed that signals from brain areas relating to speech and languages were used by the decoders, suggesting the reasonability of the machine learning method from the physiological point of view. Moreover, when calculated functional connectivity between current sources and visualized the signaling flows on brain models calculated using magnetic resonance image (MRI) and diffusion tensor images (DTI), the signaling flow was also reasonable such that signaling flow started from the auditory area when the user heard auditory stimuli and transmitted to frontal areas when the user imagined the heard vowel (Fig. 2). Using the same technique, we have recently succeeded to reconstruct the heard auditory stimuli from EEG signals (Akashi et al., under review). These results suggest the efficacy of the EEG cortical current source, and recently I have found that applying network analysis methods to EEG cortical current sources further enhanced decoding performance (Scientific Reports, 2017).

​