Zero-shot imagined speech decoding via MEG-to-listening mapping

Decoding imagined speech from non-invasive brain recordings has long been a challenge due to scarce, poorly aligned datasets. In a new paper (arXiv:2605.08075), researchers Maryam Maghsoudi and Shihab Shamma introduce a zero-shot method that sidesteps the need for large imagined speech corpora. Their key insight: leverage the richer, reliably labeled recordings of actual listening. They collected MEG (magnetoencephalography) data from trained musicians while they listened to and imagined rhythmic melodic and spoken stimuli. Using musicians improved temporal alignment between conditions. The team then built a three-stage pipeline: first, they trained six linear and neural models to map imagined MEG responses to their listened counterparts, validating against null baselines from unseen subjects to confirm stimulus-specific information was preserved. Second, they trained a contrastive word decoder exclusively on listened MEG responses, evaluating embeddings across semantic, acoustic, and phonetic representations. Finally, they applied the mapping to imagined MEG from held-out subjects, feeding the predicted-listening responses into the listened decoder.

Results show that imagined words are decodable significantly above chance using rank-based analysis, even though the decoder never saw imagined data during training. This zero-shot capability is a major step for brain-computer interfaces (BCIs), as it eliminates the need for extensive, subject-specific imagined speech training sessions. The method also scales with training data size, hinting at practical, real-world BCIs for communication. While still a proof of concept, the approach opens doors to silent speech interfaces that require no explicit imagination calibration, potentially helping paralyzed or locked-in patients communicate naturally. The paper is available on arXiv under a pending DOI.