NARRATIVE SENTENCE DECODING USING A MEG-BASED BRAIN-COMPUTER INTERFACE

Millions of individuals develop motor speech impairments due to neurological disease or injury, and restoring communication has the potential to substantially improve their quality of life. Although invasive speech brain-computer interfaces (BCIs) can achieve high decoding performance, surgical risks and limited long-term usability concerns motivate the development of scalable, non-invasive alternatives. Magnetoencephalography (MEG) provides millisecond-scale measurements of cortical dynamics and has recently demonstrated promise for decoding continuous speech perception. However, MEG decoding remains underutilized. We developed a MEG-based pipeline to decode perceived speech during naturalistic story listening (Figure 1). We analyzed approximately 10 hours of MEG data from each of two healthy English speakers while they listened to audiobooks. Sensor-level band-limited power features were extracted across five canonical frequency bands (theta, alpha, low-beta, high-beta, and low-gamma). A gated recurrent unit (GRU) decoder, enhanced with variational autoencoder-based data augmentation, was trained on these features to predict phoneme sequences drawn from an inventory of 39 phonemes and a silent character. Decoded phoneme strings were subsequently mapped to candidate sentences using a language model, with GPT-2 employed for final rescoring. Across frequency bands, low-beta power yielded the strongest performance, achieving a phoneme error rate (PER) of 62.2% and a character error rate (CER) of 65.8% on reconstructed sentences, substantially outperforming chance-level error rates (97.5% PER; 96.3% CER). These results establish a baseline for sentence-level decoding from MEG oscillatory features and support MEG, particularly in conjunction with emerging wearable optically pumped magnetometer (OPM) systems, as a promising non-invasive platform for advancing speech BCI technologies.