The past few years have seen remarkable progress in the decoding of speech from brain activity, primarily driven by large single-subject datasets. However, due to individual variation, such as anatomy, and differences in task design and scanning hardware, leveraging data across subjects and datasets remains challenging. In turn, the field has not benefited from the growing number of open neural data repositories to exploit large-scale deep learning. To address this, we develop neuroscience-informed self-supervised objectives, together with an architecture, for learning from heterogeneous brain recordings. Scaling to nearly 400 hours of MEG data and 900 subjects, our approach shows generalisation across participants, datasets, tasks, and even to novel subjects. It achieves improvements of 15-27% over state-of-the-art models and matches surgical decoding performance with non-invasive data. These advances unlock the potential for scaling speech decoding models beyond the current frontier.

Although more and more data from participants recorded with safe brain imaging devices are appearing publicly on the internet, we do not have methods that can effectively train AI on all of this data together because each data source uses different methods to collect their data and provide different labels. We built a method that resolves the differences between these data sources and is able to work without labels, allowing all of the brain imaging data on the internet to be used to train AI. In turn, this unblocks the path to building better AI-driven brain-computer interfaces by collecting and combining even more data. We hope to eventually use these methods to help paralysed patients communicate through AI that deciphers their intended speech from their brain signals.