Speech brain implants for paralysis: what the trials actually show

Brain-computer interfaces that decode attempted speech have reached the point where an implanted array of electrodes can turn the neural activity of someone who cannot speak into text on a screen, in some cases fast enough to approach conversation. In one published study a participant with ALS produced text at 62 words a minute [s1]; in another, a different implant reached 99.6% accuracy on a small vocabulary on its first day of use [s2]. These are real, peer-reviewed results and a genuine advance for people locked out of communication — and they come from single individuals, require open brain surgery, and remain experimental research rather than an approved treatment anyone can receive.

What a speech neuroprosthesis is

Conditions such as amyotrophic lateral sclerosis (ALS) and brainstem stroke can leave a person cognitively intact but unable to move the muscles of speech. A speech brain-computer interface (BCI) bypasses those muscles: surgeons place microelectrode arrays into the region of cortex that controls speech, the electrodes record the firing of individual neurons as the person attempts to talk, and a machine-learning model translates that activity into words [s1]. The output can be shown as text or spoken aloud by synthetic voice. This is not the scalp-electrode "neurofeedback" sold to consumers, which reads only smeared, low-resolution signals through the skull; the performance below depends on electrodes implanted in the brain, and the comparison with consumer EEG headbands, whose clinical claims are thin, is worth keeping in mind.

What the published results show

In a 2023 study, a participant with ALS who could no longer speak intelligibly used intracortical microelectrode arrays to drive a speech-to-text system [s1]. The BCI achieved a 9.1% word error rate on a 50-word vocabulary — 2.7 times fewer errors than the previous best speech BCI — and 23.8% on a 125,000-word vocabulary, which the authors described as the first successful demonstration of large-vocabulary decoding [s1]. Attempted speech was decoded at 62 words per minute, 3.4 times as fast as the prior record and moving toward the roughly 160 words per minute of natural conversation [s1]. The jump in the error rate from the small vocabulary to the large one is the honest headline: usable for a constrained set of words, still error-prone for open-ended speech.

A 2024 study reported a different system in a 45-year-old man with ALS, tetraparesis and severe dysarthria, who had four microelectrode arrays — 256 electrodes — implanted in his left precentral gyrus five years after his illness began [s2]. On the first day of use, 25 days after surgery, the neuroprosthesis reached 99.6% accuracy with a 50-word vocabulary after just 30 minutes of calibration [s2]. On the second day, after 1.4 more hours of training, it hit 90.2% accuracy on a 125,000-word vocabulary, and with further data it sustained 97.5% accuracy over 8.4 months, letting the man hold self-paced conversations at about 32 words per minute across more than 248 cumulative hours of use [s2]. The advance there is less raw speed than how quickly the system became usable and how well it held up.

Why sober framing matters

The temptation is to read these numbers as a solved problem or an imminent product. They are neither, for reasons the papers themselves make plain. Each result comes from a single participant [s1][s2]; performance in one person's brain does not establish how the technology will work across many, with different anatomy, disease and electrode placement. Both required neurosurgical implantation of arrays into the cortex — an invasive, irreversible procedure with its own risks — and both were conducted under investigational research protocols, not as approved medical devices available in clinics.

Durability and generalisation are still open questions: one study followed a participant for 8.4 months [s2], which is encouraging but short against a lifetime, and implanted electrodes are known to degrade over years. Accuracy also remains far from perfect once the vocabulary opens up — a 23.8% word error rate means nearly one word in four is wrong [s1]. None of this diminishes what has been achieved; restoring any conversational communication to someone who had lost it is a profound result. But it belongs in the same evidence-first frame as any other neurotechnology, from implanted brain stimulation for depression onward: report what the trials measured, in whom, for how long — and resist the leap from a striking proof of concept in one person to a promise for everyone.

Sources

• [s1] A high-performance speech neuroprosthesis — Nature, 2023-08-23. https://doi.org/10.1038/s41586-023-06377-x
• [s2] An Accurate and Rapidly Calibrating Speech Neuroprosthesis — New England Journal of Medicine, 2024-08-01. https://doi.org/10.1056/NEJMoa2314132