70.7 Correct Words a Minute. The Trial Scores a Monthly Maximum.

70.65.

Multiply 75, the words per minute in Neuralink's October 28, 2026 post, by 0.942, the share left after a 5.8% word error rate, and the product is 70.65 correct words a minute, which rounds to 70.7. One participant, the post says, hit 75 words per minute at that error rate on a 125,000-word vocabulary in the first week, and Nader Pouratian presented it the same day at the Northeast ALS Consortium, whose proceedings sit in Muscle & Nerve, doi 10.1002/mus.70436. This article did not extract that abstract, so 75 and 5.8% remain the company's figures for a talk.

A maximum is the score they wrote down

VOICE, NCT07224256, Neuralink's early feasibility study of the N1 implant and the R1 robot, is recruiting at UT Southwestern, aiming for six adults with severe speech loss and impaired arms. Primary outcomes are the shares of participants with device-related and procedure-related adverse events at 12 months. Speech is secondary, and the registry names it "monthly maximum correct word per minute." A maximum is not a mean. A first-week point can be exactly the sort of number that endpoint stores, and neither one is the average of a week of ordinary talk.

Willett and colleagues, in Nature in 2023, decoded attempted speech from participant T12 at 62 words per minute across five vocalized evaluation days, with a 23.8% word error rate on that same 125,000-word vocabulary, while silent days held the rate at 62 and moved the error rate to 24.7%. Multiply 62 by 0.762 and the vocalized correct-word rate is 47.244, so 70.65 divided by 47.244 is 1.50, the week-one report against that multi-day average, if 75 is the attempt rate and 5.8% removes words the way a word error rate usually does. Their paper sets natural conversation at 160 words per minute and cites Yuan, Liberman, and Cieri (2006): 75 divided by 160 is 46.9% of that pace, and 70.65 divided by 160 leaves 44.2% once the errors are out.

What a counted home record actually contains

Card and colleagues, in the New England Journal of Medicine in 2024, followed T15, a man with ALS and four 64-electrode arrays (256 sites) in the left ventral precentral gyrus. Copy-task error in the final five sessions was 2.5% (95% CI 2.0 to 3.1). Conversation on 925 known sentences averaged 3.7% (95% CI 3.3 to 4.3) at about 32 words per minute across 248.3 hours, and 32 times 0.963 is 30.8 correct words a minute if that approximate 32 is taken as exact. Before the implant, expert listeners got 6.8 correct words a minute from his speech and a head mouse managed 6.3. The paper puts smartphone dictation near 5% errors, a microphone benchmark the implant does not get to use, and 5.8% still sits above his 2.5% copy-task errors.

On June 15, 2026, the home follow-up in Nature Medicine says T15 produced 183,060 sentences and 1,960,163 words at 56 words per minute, inside more than 3,800 hours with no researchers present, under BrainGate2, NCT00912041. He marked 53.3% of sentences fully correct, corrected 12.9%, and called 26.1% mostly correct. Vocalized copy tasks exceeded 99% accuracy at 30.6 words per minute. Silent copy reached 96.5% at 49.7.

Divide 1,960,163 by 56 and the speaking time implied by that average is 35,002.9 minutes, or 583.4 hours. Divide 583.4 by 3,800, treating 3,800 as the floor printed beside the word "more," and speech at the stated average fills at most 15.4% of the hours the system was on. Multiply 56 by 3,800 by 60 and the naive product is 12,768,000 words, 6.51 times the 1,960,163 words the paper counts, because a rate describes minutes of talking and a duration describes a powered cart. Different clocks. No word count and no session length have been published for the first week, so 75 times 7 times 24 times 60, equaling 756,000 words, is the same kind of false product, a rate left running through a calendar the post does not claim.

Seventeen words, then a miss

On the silent copy task, 49.7 times 0.965 is 47.96 correct words a minute, a hair above Willett's 47.24, and 70.65 divided by 47.96 is 1.47, so the week-one report is faster than that prompted average and not in a different league. Words between errors are the reciprocal of the error rate, 1 / 0.058 = 17.2, 1 / 0.035 = 28.6, 1 / 0.025 = 40, and 1 / 0.238 = 4.2, which means a 12-word sentence expects 0.70, 0.42, 0.30, and 2.86 errors on those four rows, and takes 9.6 seconds at 75 per minute, 14.5 seconds at 49.7, 22.5 seconds at 32, and 11.6 seconds at 62.

Metzger and colleagues, in Nature in 2023, reported a median 78 words per minute and a median 25% word error rate on silent sentences from a 1,024-word set. Then 78 times 0.75 is 58.5 correct words a minute, and 125,000 divided by 1,024 is 122.1, so the faster raw speed sits on a list 122 times smaller. A 2024 review and its supplement table print 78.3 words per minute at 25.5% errors (58.3 correct words) and put Moses and colleagues' 2021 spelling prosthesis at 15.2 words per minute with 25.6% errors on 50 words. Card's journal sentence cites that same 25.5% on 1,024 words. Half a point does not change the column.

One other reading stays open, because the registry already names the endpoint in correct words per minute: if the posted 75 is that maximum, multiplying by 0.942 counts the errors twice, and the attempt rate implied by 5.8% errors is 75 / 0.942 = 79.6 words a minute. The post's grammar, a rate "at" an error rate, matches the way Willett prints 62 and 23.8% as separate measurements, and that is the reading in the table, still a reading rather than a decoded methods section.

What You Can Do

If you counsel someone with ALS, set the week-one point beside T15's audited home total and say which number is a maximum, then ask any later update for a mean, a sentence count, and a word error rate scored as the New England Journal paper defines it, insertions, deletions, and substitutions against a known sentence. If you write a headline, do not multiply 75 by the hours in a week, and do not multiply 56 by 3,800. Watch the monthly maximum on NCT07224256, the 12-month adverse-event proportions, and whether a full paper replaces the post. Six planned implants against the CDC National ALS Registry dashboard's 2026 projection of 34,720 living adults is 34,720 / 6 = 5,787 projected cases per trial slot, a capacity ratio on a projection row and not a forecast of who is offered a device.

Limitations

This analysis did not read the abstract, see the sentences, or re-score a recording. The post does not say whether 75 is a peak, a session mean, or the monthly maximum, whether sentences were prompted, how many were scored, or whether 5.8% counts insertions and deletions. Expecting 0.70 errors in 12 words treats errors as independent, and a language model can bunch them. Because 32 words per minute is approximate, 30.8 inherits that, and because 3,800 hours is a floor, 15.4% is a ceiling on the speaking share. The CDC page printing 34,720 dates its statistics note to September 1, 2023, calls calendar-year 2018 the latest detailed prevalence, and was reviewed on June 25, 2026. Its sex rows, 17,583 and 14,977, add to 32,560, not 34,720. No price, operative time, or adverse-event count for VOICE is public here, so nothing in this piece is a bill.

The Strongest Case Against This Analysis

A fair objection says 70.65 is the wrong lead, and it should be stated at full strength. Card's silent copy task already delivers 47.96 correct words a minute at 3.5% errors, across sessions in a paper, from a man who then sent 1,960,163 words anyone can divide. A company post about one unnamed person in week one can be the best hour of an endpoint that stores a maximum. Willett showed offline that a stronger language model alone moved 23.8% errors to 17.4%, and closer-in-time test sentences moved that to 11.8%. Moses and colleagues, on a 50-word list in the same journal in 2021, went from a 60.5% word error rate with no language model to 25.6% with one, at a median 15.2 words per minute, inside NCT03698149, and they called speech decoding generally usable below 30% errors, a line that does not separate 5.8% from 2.5%. Willett's speech paper treats a 2021 handwriting implant as the old 18-word record. That implant reported 90 characters per minute at 94.1% raw online accuracy. If 5.8% is a prompted best case and 3.5% is an average, the week-one report is faster and not cleaner. Grant that. The 1.47-fold gap does not retire the objection.

The Bottom Line

On the reading that 75 is an attempt rate, the week-one report yields 70.65 correct words a minute, 1.50 times Willett's vocalized average and 1.47 times Card's silent-copy average, with an error expected about every 17 words and about 0.70 errors inside a 12-word sentence that takes 9.6 seconds to emit. The registry scores a monthly maximum of correct words per minute, the primary results are 12-month adverse-event proportions on a planned six people, and the only home total anyone can divide, 1,960,163 words inside a 3,800-hour floor, fills at most 583.4 hours at 56 words a minute. Carry the correct-word arithmetic, and leave the calendar product, plus any claim that one reported week has beaten 183,060 counted sentences, until a mean, a sentence count, and a known text are in a paper.

Related

• A Brain Implant Read Speech and Gestures at Once. The 56-Condition Wall Is the Real Story.
• 421 Electrodes Just Outperformed 1,024. The Brain-Computer Interface Industry Is Optimizing the Wrong Variable.
• Neuralink Just Deleted Brain Surgery's Most Delicate Step. The Real Question Is Whether That Unlocks 10,000 Implants a Year or 50.