Brain2Qwerty v2 Achieves Real-time Brain-to-Text Decoding

A new milestone in non-invasive brain-to-text decoding has been announced with the release of Brain2Qwerty v2. This advanced system builds upon its predecessor, which was recently published in Nature, by offering the highest-performing end-to-end pipeline for real-time sentence decoding directly from raw brain signals. Brain2Qwerty v2 moves beyond character-level performance to decode full words and their semantics, significantly enhancing overall communication accuracy. The research involved training the system on approximately 22,000 sentences from nine volunteers, each recorded for ten hours using MEG devices while typing. The system leverages end-to-end deep learning on raw neural data and fine-tuned large language models to bridge the gap between noisy brain signals and coherent language. Promising results include an average word accuracy of 61% across participants, with the top performer achieving 78% word accuracy and decoding over 50% of sentences with one or fewer word errors. To foster further advancements, the full training code for both v1 and v2, along with the v1 dataset, has been made publicly available.