Neuralink's implant hit 75 words a minute, but there's no paper yet

Seventy-five words a minute, and one mistake every 17 words. That is the result Neuralink posted on X on October 28, 2026: a person with amyotrophic lateral sclerosis (ALS) dictating text to a brain implant in the very first week after it was switched on. Three years ago, the best system of this kind got roughly every fourth word wrong.

If the number holds up, it is the best start any speech implant has reported: no system has paired this speed with this accuracy in its first week. It sounds like a record. For now, though, the whole announcement fits in two sentences. They say nothing about how many people took part, how the test was run, or how often the participant can actually talk that way. The best benchmark already exists: a man with ALS who has spent almost two years talking through an implant at home, and has gone back to work.

Two sentences instead of a paper

In essence, the post says this. The company is enabling “speech to text and voice at conversational rates” in people with ALS. One participant reached 75 words per minute at a 5.8% word error rate on a 125,000-word vocabulary in their first week. The same day, neurosurgeon Nader Pouratian presented the data at the annual meeting of NEALS, a consortium of ALS researchers.

Pouratian leads the VOICE trial. According to its registry entry , it is an early study at a single hospital in Dallas, with six participants. Its main goal is safety: complications and side effects in the year after surgery. Speech is a secondary goal, and it is measured as the best rate reached each month.

That raises the first question about the post. If the trial counts speed by its peak, 75 words per minute may well be a best moment rather than an average. The post leaves something else out too: whether the participant was reading sentences off a screen or saying whatever came to mind. For speech implants the gap is huge, because repeating a prompt is far easier than holding a conversation.

How the brain dictates

Someone who can barely speak and has limited use of their hands (because of ALS, a stroke or a spinal cord injury) is left typing with a finger or with their eyes. It is slow and exhausting. Hence the interest in a direct line from the brain.

A speech implant doesn’t read thoughts in the everyday sense. Its electrodes sit in the part of the cortex that drives the lips, tongue and larynx. When a person tries to say a word, the neurons there fire much as they did before the illness, even if the muscles no longer respond. Software learns to recognize speech sounds in those bursts, and a language model assembles the sounds into plausible words. Think of it as phone autocorrect, except the keys are being pressed by the cortex.

In September, Neuralink described a participant named Terry. He first trained the algorithm by miming words, then simply thought them and heard his old voice come out of the speaker. His N1 implant sits in the skull, recharges wirelessly and sends its data by radio. Neuralink doesn’t say whether the new numbers are Terry’s.

Three years of the race in numbers

In August 2023, Nature ran two papers in the same issue that kicked off the current race. Teams at Stanford and San Francisco chased speed and got close to half the pace of normal speech, but paid for it in accuracy. The Stanford authors set the bar themselves: ordinary conversation runs at about 160 words per minute.

Then accuracy took the lead. In 2024, a UC Davis team described in NEJM a 45-year-old man with ALS, Casey Harrell. On day two his implant understood 90.2% of words, or about one error in ten. Neuralink claims one error in 17 in the first week, at more than twice the speed Harrell managed in free conversation back then, though its test conditions are unknown.

Casey Harrell’s 3,800 hours

The most useful benchmark arrived in June 2026. Nature Medicine published results from almost two years in which Harrell used his implant at home for more than 3,800 hours. He produced nearly 2 million words and more than 183,000 sentences, at an average of 56 words per minute. His home care team runs the system without researchers, and Harrell has gone back to full-time work as an environmental advocate.

The same study shows how lab accuracy drifts from real life. In controlled tests the system recognized more than 99% of words correctly. In everyday speech, Harrell himself rated 92% of sentences as accurate or mostly accurate. Roughly one sentence in twelve came out with a noticeable error. That is something you feel in daily conversation, even though by 2023 standards it would have looked like a miracle.

Harrell’s setup has its own limits: his arrays are made by another company, Blackrock Neurotech, and he is still wired to an external computer. Neuralink’s wireless design may well prove easier to live with at home. But the only way to find out is the same one: months of ordinary life, not a week.

How far from a real conversation

Even at its best, 75 words per minute is less than half the pace of ordinary speech. That’s enough to dictate an email. Not enough to argue over dinner.

Conversation also depends on how fast you can reply: a pause of a few seconds breaks the rhythm, and the other person finishes your sentence or drifts off to their phone. Neuralink said nothing about latency, which matters more for a live chat than a record speed.

In three years, speech implants went from one word in four being wrong to fewer than one in a hundred in tests. Harrell says no technology he has tried let him communicate so close to his natural way of speaking. The VOICE participants have yet to show this not in their first week, but at home, month after month.

It took Harrell almost two years to reach his 2 million words. How many words will a Neuralink participant have logged by then, and how many of them will be wrong?