The system learns directly from raw neural signals rather than relying on a hand-built event-detection pipeline. Meta trained v2 on approximately 22,000 sentences from nine volunteers, used large language model fine-tuning to add semantic context, and reports a 61% word accuracy overall with a best-participant result of 78%.
Brain2Qwerty is useful for neuroscience, non-invasive brain-computer interface research, and assistive-communication investigations. Meta is releasing the full training code for v1 and v2, while the BCBL partner is releasing the v1 dataset, giving researchers artifacts for reproduction and for studying how neural recordings can be mapped to coherent language.
