Meta Brain2Qwerty v2: non-invasive typing at 39% WER

Meta's FAIR research team has released Brain2Qwerty v2, a model that reconstructs full sentences from non-invasive magnetoencephalography recordings, achieving an average word error rate of 39 percent and a best-participant WER of 22 percent. The study used nine healthy volunteers, each recorded for ten hours, producing a dataset of 22,000 typed sentences.

How does Brain2Qwerty v2 work?

Brain2Qwerty v2 decodes typed sentences from MEG by processing brain signals at three levels—characters, words, and full sentences—then fine-tuning a language model to convert noisy encoder outputs into coherent text. The team replaced earlier hand-built recognition steps with deep learning, and used Qwen3 as the sentence-level language model to shape outputs.

Researchers captured measurable activity primarily from the motor cortex while participants heard a sentence, paused, and then typed it without seeing the text on screen. Version 2 abandons the requirement for exact keystroke timestamps by operating on a continuous signal window that assigns characters asynchronously, which the team says removes a barrier to real-time use even though real-time capability is not yet achieved.

How does it compare with the original and other baselines?

On word- and meaning-level metrics Brain2Qwerty v2 outperforms simpler methods: the model reaches a 39 percent word error rate, versus 55 percent for the raw encoder and 43 percent for the N-gram approach used in Brain2Qwerty v1. At the same time v2 produces more character-level errors, with a 31 percent character error rate compared to 28 percent for the raw encoder and 26 percent for the N-gram model.

The language model explains the trade-off: Qwen3 nudges outputs toward fluent sentences, which improves word- and semantic-level measures but can invent grammatically plausible yet incorrect sentences when the brain signal is ambiguous. For the best participant, 28 percent of sentences were decoded perfectly and 47 percent contained at most one wrong word. For context, invasive implanted interfaces still perform far better, achieving below two percent word error rate for typing.

What role did data and automation play in the improvement?

The team credits about ten times more recordings per person and much more varied sentences for enabling the asynchronous decoding approach and higher-level modeling. They also used an auto-research component: three independent agents based on Claude Opus 4.6 were given optimization tasks and discovered techniques such as label smoothing, modality dropout, and shorter prompts that consistently reduced error rates and beat a standard optimization method.

Those agents had limits. When given open-ended code changes they crashed most compute jobs, so human researchers remained essential. The paper also notes practical steps toward clinical feasibility: tests with portable MEG sensors showed that even half the sensors can deliver nearly full performance.

Why it matters

Brain2Qwerty v2 narrows the gap between non-invasive and invasive brain-computer interfaces by demonstrating coherent sentence reconstruction without surgical implants. The shift from keystroke-aligned decoding to an asynchronous, sentence-level approach and the use of a modern language model yields a measurable gain in word- and semantic-level communication, which matters because successful communication depends more on meaning than exact characters.

The study also ties into a broader FAIR research program: earlier work decoded perceived speech from MEG and EEG and generated images from brain activity, and the team positions these projects as tools for both assistive technology and basic neuroscience.

What to watch

Look for results from more diverse participant sets, experiments that remove the requirement for actual finger movements, and any demonstrations of real-time decoding. The team highlights collecting more recordings as a straightforward way to improve accuracy, and portable room-temperature MEG sensors as the plausible next hardware step toward clinical trials.