Brain2Qwerty: MEG decoding of QWERTY typing (Lévy et al., 2025)
Measured by Lévy, Zhang, Pinet, Rapin, Banville, d'Ascoli, King et al. · arXiv:2502.17480 (2025)
Inputs
The measured or assumed values behind the calculations, each with its source.
- rate = 152 char/min
- Mean typing speed of the able-bodied participants on a standard QWERTY keyboard: 152.0 ± 3.2 char/min (mean sentence-production time 5.7 ± 0.2 s). This is the keyboard's rate, set by the participants' own hands; the brain decoder reconstructs the already-typed text offline rather than producing it.
- H = 1.0 bits/char
- Text entropy (Shannon); the same ~1 bit/char convention applied to QWERTY, eye-typing and the other text-entry entries. Stimuli are declarative Spanish sentences in upper case without accents.
- P = 0.68
- Per-character accuracy = 1 − CER. With MEG, Brain2Qwerty's character error rate is 32 ± 0.6% across 35 subjects (best participant 19%). EEG is far worse at CER 67% (per-character accuracy ≈ 0.33), reflecting its lower signal-to-noise ratio.
- N = 29
- Decoder output classes: 26 Latin letters + space + a numbers class + an other-special class = 29. Bounds the raw key channel under a uniform prior (log2(29) ≈ 4.86 bits/key).
- T_key = 0.395 s/key
- Key-selection interval for the Wolpaw bound: 60 / 152 char/min = 0.395 s/key.
Strictest ITR
Each scoring method is an upper bound on the channel, so the headline is the strictest (smallest) one for this entry. Use the score selector on the home page to view any single method across entries.
-
Correct characters per minute
152 char/min × (1 − 0.32 CER) = 103.4 correct char/min
The keyboard rate (152 char/min) discounted by the MEG average character error rate (32%): the rate at which correctly-decoded text is delivered.
-
Bits per character
H ≈ 1.0 bit/char (Shannon)
-
Information transfer rate
103.4 char/min × 1.0 bit/char ÷ 60 s/min = 1.72 bits/s
What counts as a bit depends on the action space. The number of distinguishable actions and how likely each one is are design choices of the task, not the sensing hardware. The same modality can present a fixed set of targets, a set pruned per step by a grammar or language model, or a continuous control space. Each of these changes how many actions are live and how the probability mass is spread, and therefore the information per selection. Read the action space below before comparing headline numbers across entries.
Action space
What the user can produce at each step, and how those options are distributed.
- Structure
- Fixed set of targets
- Size
- 29 distinguishable actions
- Prior
- Context-conditioned: likelihoods depend on prior actions
- Notes
- The effector is a regular QWERTY keyboard: able-bodied participants physically typed briefly-memorized sentences with their hands, and Brain2Qwerty reconstructs that text from non-invasive MEG. It is an offline decoding result, not a closed-loop interface: the typed keystrokes are the ground-truth labels (used for both training and scoring), the model is non-causal and operates at the sentence level (no real time), and a sentence-level language model corrects the character predictions. Because the keystrokes come from intact hand motor function, the 152 char/min rate is the keyboard's, not a volitional neural channel a non-communicating patient could drive — the authors flag exactly this as the open problem. The sentence set is also small and closed (128 unique Spanish sentences, 5–8 words, split 80/10/10). The headline figure is the strict character-entropy throughput net of the 32% CER; the Wolpaw key-channel figure is the uniform-prior upper bound on the same channel and runs higher. Listed as a non-invasive frontier datapoint, with the keyboard effector made explicit so it is not read as a speller.
Comparability The strictest bound here is the Shannon entropy of the output text, under one predictor held constant across the whole atlas (≈1 bit per character). That shared predictor makes it directly comparable to every other text entry (keyboards, spellers, silent speech and speech BCIs) regardless of prior or vocabulary size. For most text interfaces it comes out tighter than the raw-selection bounds, but not always. Where a small vocabulary makes Wolpaw tighter, that wins instead. Any Fitts, Wolpaw or log₂(N) figure shown below is another bound on the same channel. Switch the home-page score selector to compare one across entries.
Other bounds considered for the headline
Also valid upper bounds for this entry and eligible to be the headline. They just came out looser than the strictest above. Pick any of these in the home-page score selector.
-
Bits per selection (Wolpaw formula)
B = log2(N) + P*log2(P) + (1-P)*log2((1-P)/(N-1)) = log2(29) + 0.68*log2(0.68) + 0.32*log2(0.32/28) = 2.415 bits / selection
Term 1 is the information if every choice were correct; terms 2-3 subtract the bits lost to the error rate, assumed spread evenly over the other N-1 targets.
-
Selections per second
T = 0.395 s/selection -> 1 / 0.395 = 2.532 selections/s
-
Information transfer rate
ITR = B * selections/s = 2.415 * 2.532 = 6.115 bits/s
Other score types
Bounds the atlas keeps out of the default strictest headline: as-reported figures, alternate task conditions, or raw-channel ceilings that shouldn't win the headline by default. Each still carries a score type, so the home-page selector ranks this entry on it when you choose that type. Read its derivation before comparing across entries.
-
Achieved-bitrate credit per net-correct key
N = 29 → log2(N − 1) = log2(28) = 4.81 bits per net-correct selection (field-standard achieved bitrate, e.g. Webgrid; Nuyujukian 2015, which introduced the metric, used log2(N)).
-
Net-correct key rate
net-correct = 2P − 1 = 2(0.68) − 1 = 0.36 of keys. At 0.395 s/key (152 char/min) → 0.36 / 0.395 = 0.911 correct/s.
A decoding error commits the wrong character rather than timing out, so incorrect = 1 − P. Same N (29), MEG per-character accuracy (68%, i.e. 32% CER) and key interval (0.395 s) as the entry's Wolpaw calc. This is the one entry where the 2P − 1 netting bites hard: achieved (4.38) falls well below the entry's own Wolpaw bound (6.11), because Wolpaw's mutual-information term still credits ~2.4 bits/key at 68% while achieved nets each wrong key one-for-one. Both stay above the 1.72 bits/s Shannon headline, so neither ranks. At the EEG variant's 33% accuracy (67% CER) 2P − 1 goes negative and the achieved bitrate clamps to 0 — the clearest illustration of why this metric is a poor fit below ~55% accuracy and is never used for ranking.
-
Achieved bitrate
4.81 bits × 0.911 correct/s = 4.38 bits/s.
Source
- Authors
- Lévy, Zhang, Pinet, Rapin, Banville, d'Ascoli, King et al.
- Publication
- arXiv:2502.17480, 2025
- Reference
- Meta AI publication page