Researchers have released Brain2Qwerty v2, an end-to-end pipeline capable of decoding full sentences in real time from noninvasive magnetoencephalography (MEG) brain recordings. The system aims to approach levels of accuracy previously seen mainly with invasive, surgically implanted methods while avoiding surgical procedures.
What the work includes
- The team is publishing the full training code for Brain2Qwerty v1 and v2. The partner institution, the Basque Center on Cognition, Brain, and Language (BCBL), is making the v1 dataset available.
- Rather than relying on hand-crafted signal processing or event detectors, the pipeline performs end-to-end deep learning directly on raw MEG signals.
- Large language models were fine-tuned on neural data so the system can use semantic context to turn noisy brain recordings into coherent text.
- AI agents were used to explore optimizations for the decoding pipeline, with final training configurations selected manually by engineers.
Data and participants
Brain2Qwerty v2 was trained on approximately 22,000 sentences collected from nine volunteer participants. Each participant was recorded for about 10 hours while wearing a MEG device and actively typing.
Performance: accuracy and outcomes
- Overall, the system achieves a word accuracy of 61% on noninvasive recordings, a substantial improvement over the 8% word accuracy reported for other noninvasive methods.
- For the best participant, the system reached 78% word accuracy; for that individual, more than half of all decoded sentences contained one word error or fewer.
Why this matters
Invasive approaches such as stereotactic electroencephalography and electrocorticography have shown that a neuroprosthesis combined with an AI decoder can restore communication for people with severe brain injuries, but those techniques are difficult to scale. A noninvasive pipeline like Brain2Qwerty v2 could help close the gap and expand access to restorative communication technologies.
Data scaling and future directions
The researchers report a log-linear improvement in decoding accuracy as data volume increases, suggesting that additional gains — and a smaller gap with surgical approaches — may be achievable through further data collection.
This work is part of a broader effort to build open foundational models of the brain, mentioning components such as Tribev2 for perception encoding, NeuralSet for large-scale brain-data processing, and NeuralBench for systematic model evaluation. The team is also supporting community efforts through a $5 million fund under the Digital Brain Project to encourage open datasets.
The authors say that conducting this research openly may accelerate neuroscience advances to identify, diagnose, and treat neurological disorders more quickly than isolated efforts.
Availability
The team is releasing the Brain2Qwerty v1 and v2 training code, and the Basque Center on Cognition, Brain, and Language (BCBL) will release the v1 dataset to support further research and replication by the community.
