Towards decoding inner speech from EEG and MEG

Abstract

Despite the prevalence of inner speech in everyday life, research on this has been limited, particularly when it comes to non-invasive methods. This preprint aims to fill this gap by using EEG and MEG to collect data from three different inner speech paradigms, and by conducting an initial decoding analysis. Specifically, we tested silent reading, repetitive inner speech, and generative inner speech tasks. We collect a high number of inner speech trials from a few participants. Besides comparing across recording modalities we also compare across inner speech types. Our aim is to analyse the decodability of inner speech within each task and between tasks by the use of transfer learning. We find that in both EEG and MEG, silent reading can be decoded relatively well with 30-40% accuracy across 5 words. However, the decoding performance of both types of inner speech is mostly at chance level. This prohibited further transfer learning investigations between tasks. While the inner speech results are primarily negative, we believe our exploration of data size and various decoding methods is valuable. The dataset itself is useful for the research community as it contains a much larger number of trials within one participant than any other inner speech dataset. Having multiple sessions also allows for testing across-session performance. Finally, we systematically compare silent reading decoding performance within 3 participants across four non-invasive modalities. These are EEG, 2 types of MEG machines, Elekta and CTF, and optically-pumped magnetometers (OPMs). We also compare the spatiotemporal dynamics of silent reading between these modalities. This is especially aimed at validating OPMs as a new kind of non-invasive brain recording technology. We find comparable performance to EEG, but OPM performance did not reach traditional MEG.