Researchers released Brain2Qwerty v2, an updated system that translates brain activity into text. The team included contributors at Meta, the French National Centre for Scientific Research (CNRS), the Hospital Foundation Adolphe de Rothschild, the Basque Center on Cognition, Brain and Language, Paris Cité University, and the French Institute for Research in Computer Science and Automation (Inria).
How the system works
Brain2Qwerty v2 processes magnetoencephalography (MEG) recordings in three stages:
- An encoder splits brain activity into characters. The encoder is composed of a convolutional neural network (CNN) followed by a CNN/transformer hybrid called a conformer; it produces character embeddings and classifies signals into character sequences.
- An "aligner" converts character embeddings into word embeddings. The aligner groups character embeddings according to predicted spaces, averages them per word, and trains to increase similarity between the resulting word embedding and the Qwen3 embedding of the ground-truth word while decreasing similarity to other words.
- A fine-tuned language model (Qwen3-4B) corrects the decoded sequences. The authors fine-tuned Qwen3-4B for each subject using a LoRA adapter and averaged the adapters across subjects.
To train and evaluate the system, the team recorded MEG data from nine subjects typing English sentences. The dataset comprises about 90 hours of recordings and roughly 22,000 examples. MEG is a non-invasive technique that measures the brain's magnetic activity.
Results and numbers
- Brain2Qwerty v2 achieved a 39% word error rate (WER), compared to 43% WER for the original Brain2Qwerty (v1).
- The encoder's character error rate (CER) improved with more training data: at about 20 hours of data the CER was roughly 50%, while at 90 hours it fell to about 25%. The authors did not observe a performance plateau before exhausting the available data.
- Training across multiple subjects outperformed per-subject training. For the median performance across the nine subjects, training on just that individual produced a 66.5% WER, while the authors' combined training method achieved a 47.8% WER.
Background and openness
The first Brain2Qwerty study compared MEG and EEG (electroencephalography) and found that MEG enabled more accurate text prediction. Brain2Qwerty v2 uses only MEG, modifies the model architecture, and trains on more data than v1. The authors open-sourced the training code for both versions and released the data for v1.
Why this matters
The findings confirm that increasing training data improves performance and highlight that pooling data across individuals can be beneficial — a contrast to many brain–computer-interface approaches historically trained for a single user. This suggests that MEG-based text-decoding systems could improve as they are trained on more subjects, analogous to how large language models have improved with larger datasets.
While invasive approaches that implant electrodes directly into the brain have achieved single-digit error rates for communication in some patients, non-invasive MEG systems like Brain2Qwerty v2 still report substantially higher error rates. Nonetheless, incremental reductions in error rate represent meaningful progress toward non-invasive alternatives that avoid surgical risks.
Summary
Brain2Qwerty v2 shows measurable improvements in MEG-based decoding of typed text: lower WER than v1, clear gains from more data, and better performance when trained across subjects. The system is not yet as accurate as invasive methods, but the results point to continued gains with more data and model development.
