What it is
The authors trained Brain2Qwerty, a three-stage deep network, to reconstruct sentences from the brain activity of 35 healthy volunteers as they typed briefly memorized sentences on a keyboard. With magnetoencephalography, which reads the magnetic fields the cortex produces, the model reached an average character error rate of 29%, far ahead of the 65% it managed with scalp electroencephalography, across 193,000 characters recorded magnetically and 146,000 electrically. For the best participant the rate fell to 18%, low enough to decode a variety of sentences outside the training set perfectly, while the weakest reached only 47%. The architecture matters as much as the sensor: a convolutional stage, a sentence-level transformer and a character-level language model together beat EEGNet, a standard brain-computer-interface model, by 2.5-fold on magnetoencephalography, and the language model corrected even the typing mistakes the participants made. The model's confusions track the physical layout of the keyboard, which the authors read as evidence it is decoding the motor command rather than language itself.
Why it matters
Brain-to-text systems that work today need electrodes implanted in the brain, with the risks of neurosurgery and the problem of keeping an implant working for years. This decodes sentence production from outside the skull. At 29% on average and 18% for the best participant it remains well behind implants, which reach 90 characters a minute at under 6% error, but the gap is no longer categorical. The size of the sensor difference is the other result: 29% magnetically against 65% electrically suggests a safe interface may do better to prioritise magnetic over electrical sensing, and the authors point to optically pumped magnetometers, which might one day offer that sensitivity in a wearable form.
Underlined numbers link to their source. Every metric and quoted figure is listed under Sources and data below.
Filed underbrain-computer interface, MEG, deep learning, neurotechnology, non-invasive
