BCIwiki (bciwiki.com) — Language models help decode imagined speech from magnetoencephalography (MEG) signals more than they help decode heard speech, according to a study posted to arXiv on September 25, 2026. The authors are Maryam Maghsoudi, Sai Samrat Kankanala, Shihab A. Shamma and Sriram Ganapathy. The preprint has not been peer reviewed.
Decoding imagined speech is a key goal for brain-computer interfaces (BCIs), but it is limited by weak neural responses, low signal-to-noise ratio and scarce datasets, the authors wrote. Using a paired listened-imagined MEG dataset, they added language-model information at two stages. First, a contrastive neural decoder aligns MEG representations with acoustic and contextual language representations, which the authors say improved cross-subject word decoding for both listened and imagined speech. Second, at inference, a neural-constrained beam-search framework combines neural evidence with language-model next-word probabilities.
For imagined speech, the best-performing balance between neural and language-model evidence shifted toward the language model, and the gain over neural-only decoding was larger. The authors conclude that language priors are most useful when neural evidence is weaker, making them particularly valuable for imagined-speech BCIs. The abstract does not report specific accuracy figures.