The Meta AI website recently announced the introduction of Brain2Qwerty v2, an end to end deep learning pipeline capable of real time sentence decoding from non invasive brain recordings. Unlike traditional neuroprosthetics that require surgical implants, this system decodes language directly from raw neural signals captured via a magnetoencephalography device worn by participants while typing. To bridge the gap between noisy brain data and coherent text, the system fine tunes large language models on neural recordings, allowing the AI to leverage semantic context during translation.
The pipeline was trained on approximately 22000 sentences across nine volunteers and achieved a 61 percent average word accuracy rate, which jumps to 78 percent for the top performing participant. Meta noted that decoding accuracy improves log linearly with data volume, suggesting that further data scaling could narrow the performance gap with invasive surgical methods. To accelerate open neuroscience research, Meta and its partner, the Basque Center on Cognition, Brain, and Language, have open sourced the full training code alongside the initial project datasets.