Neuralink VOICE Speech BCI (Pouratian et al., 2026)
Measured by Pouratian, Mills, Mupambo, Granatir, Albrecht & Amadio · Muscle & Nerve: Proceedings of the 25th Annual Meeting of the Northeast ALS Consortium (abstract) (2026)
Inputs
The measured or assumed values behind the calculations, each with its source.
- N = 125000
- 125,000-word open vocabulary used in the structured open-vocabulary evaluation (Results)
- rate = 75 word/min
- Participant 3 (first-in-human 3,072-channel N1 array), reached in under a week post-implant (Results). The abstract does not say whether this is a peak or a session average. Participant 2 (1,020 channels) reached 54 wpm on the same evaluation.
- P = 0.942
- 1 − 5.8% word error rate for participant 3 on the 125,000-word evaluation (Results). Participant 2: 6.6% WER. Participant 1: 45 wpm at 7% WER, but on a 220-word vocabulary, so not comparable.
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.
-
Error-corrected words per minute
(1 − WER) × rate = 0.942 × 75 = 70.7 net word/min
Participant 3's open-vocabulary operating point. Participant 2's 54 wpm at 6.6% WER gives 50.4 net word/min → 4.20 bits/s on the same basis.
-
Shannon per-word entropy of English
H ≈ 5.0 bits/word
Credits only the information in the English produced, independent of vocabulary size.
-
Information transfer rate
70.7 word/min × 5.0 bits/word ÷ 60 s/min = 5.89 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
- Context-dependent (the live set changes per step)
- Size
- 125000 distinguishable actions
- Prior
- Context-conditioned: likelihoods depend on prior actions
- Notes
- First-in-human VOICE trial of the fully implanted, wireless Neuralink N1 decoding attempted speech from left-hemisphere speech motor cortex (ventral premotor and adjacent speech/language regions) in three adults with ALS. Same large-vocabulary, language-model-mediated structure as the other free-speech BCIs: 125k words nominally, but the decoder reweights candidates by context (the abstract also describes a multi-participant foundation model and conversational/personal-context conditioning), so the uniform-prior Wolpaw figure is not directly comparable to the atlas English-output convention. Source is a conference abstract: no methods detail on the evaluation task (prompted vs. free), the number of sentences, or whether 75 wpm / 5.8% WER is a best session or a sustained average, so this is an early operating point rather than a peer-reviewed sustained rate.
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 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.
-
Bits per selection (Wolpaw formula)
B = log2(N) + P*log2(P) + (1-P)*log2((1-P)/(N-1)) = log2(125000) + 0.942*log2(0.942) + 0.058*log2(0.058/124999) = 15.63 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 = 0.8 s/selection -> 1 / 0.8 = 1.25 selections/s
-
Information transfer rate
ITR = B * selections/s = 15.63 * 1.25 = 19.538 bits/s
-
Achieved-bitrate credit per net-correct word
N = 125,000 → log2(N − 1) = log2(124,999) ≈ 16.93 bits per net-correct selection (Nuyujukian 2015, which introduced the metric, used log2(N); at this N the difference is negligible).
-
Net-correct word rate
net-correct = 2P − 1 = 2(0.942) − 1 = 0.884 of words. At 75 word/min (0.8 s/word) → 0.884 × 75 / 60 = 1.105 correct/s.
A word error commits the wrong word rather than timing out, so incorrect = 1 − P. Same N (125,000), word accuracy (94.2%) and rate (75 wpm) as the entry's Wolpaw calc. As that calc's own caveat notes, the 125k-word vocabulary is context-reweighted by the language model each step, so feeding it into log2(N − 1) is a large-vocabulary capacity view (~17 bits/word) — raw channel capacity, not communication, analogous to the nagel code-space figure. The ranked communication rate is the 5.89 bits/s word-entropy Shannon.
-
Achieved bitrate
16.93 bits × 1.105 correct/s = 18.7 bits/s.
Source
- Authors
- Pouratian, Mills, Mupambo, Granatir, Albrecht & Amadio
- Publication
- Muscle & Nerve: Proceedings of the 25th Annual Meeting of the Northeast ALS Consortium (abstract), 2026
- Paper
- 10.1002/mus.70436