Meta unveils AI system that translates brain signals into text with 61% accuracy
Brain2Qwerty v2 uses non-invasive brain scanning and deep learning to decode typing-related neural activity, a massive jump from the 8% accuracy of previous methods.
Meta has released Brain2Qwerty v2, an artificial intelligence system that converts noninvasive brain recordings into text in real time without requiring a surgical implant.
The system achieved an average word accuracy of 61%, compared with about 8% for previous noninvasive methods. Accuracy reached 78% for the strongest participant, with more than half of that person’s sentences decoded with no more than one incorrect word.
Meta trained the model on roughly 22,000 sentences collected from nine healthy volunteers. Each participant spent about 10 hours inside a magnetoencephalography scanner while typing sentences, giving the model around 10 times more training data per person than its previous version.
Brain2Qwerty v2 processes continuous brain activity through separate components designed to detect characters, align words and reconstruct complete sentences. Meta also fine tuned large language models on neural data to help the system use semantic context when interpreting noisy signals.
The new version removes a major limitation of the original system, which required researchers to know the timing of each key press in advance. Brain2Qwerty v2 instead generates sentences directly from continuous recordings, allowing the pipeline to operate in real time.
The research does not represent unrestricted thought reading. Participants were healthy volunteers who actively typed memorized sentences, and the system relied on a large MEG scanner that remains impractical for most clinical settings. Meta also acknowledged that the current error rate is still too high for everyday communication.
Meta said performance improved as the amount of training data increased, without showing a clear plateau. The company believes additional data and advances in wearable MEG sensors could narrow the remaining gap with brain computer interfaces that require surgical implants.
The company released the full training code for both versions of Brain2Qwerty, while the Basque Center on Cognition, Brain and Language published the dataset used to develop the first version. Data from the second version remains unavailable pending journal publication.