For decades, the ceiling for someone who lost speech to ALS or a brainstem stroke was an eye-tracking speller: stare at letters on a screen, blink to confirm, build a sentence one glance at a time. It works, and it is still the only clinically available option most patients have today. But a wave of brain-computer interface research is now decoding attempted speech directly from the motor cortex, and 2026 has produced enough real clinical data to actually compare the approaches instead of just marveling at the demo reels. Three groups in particular are worth tracking closely, and they disagree, sharply, on how many electrodes you need and how invasive it is fair to get to answer that question.
Reading attempted speech straight from the cortex
The most clinically mature result in this space comes out of the academic BrainGate2 consortium, specifically a UC Davis team led by neuroscientists Sergey Stavisky and David Brandman. In a study published in the New England Journal of Medicine in August 2024, the team implanted four microelectrode arrays, 64 microelectrodes each, for a total of 256 channels, along the left precentral gyrus of a 45-year-old participant named Casey Harrell, whose ALS had left him with severely slurred speech. Twenty-five days after surgery, researchers turned the system on. It decoded his attempted speech into text and synthesized voice at more than 99 percent accuracy on an initial 50-word vocabulary, and it held up well as the vocabulary expanded over the following months, letting him communicate at roughly 32 words per minute.
What makes this particular result stand out isn’t the accuracy number, which other teams have matched. It’s where they tested it next. A follow-up study published in Nature Medicine in June 2026 moved the system out of the lab and into Harrell’s actual home, checking whether the decoding held up without a room full of engineers nearby. That is a meaningfully different bar than a single showcase session, and it is the kind of unglamorous, longitudinal evidence this field badly needs. Worth being precise about where this stands, though:
The BrainGate2 trial is still an investigational device study and is still actively enrolling participants
UC Davis’s own communications about the work carry an explicit caution: this is an investigational device, not a marketed product
The 256-electrode array requires open craniotomy to implant, the most invasive option of the three covered here
Accuracy figures come from a single participant’s data over months, not a large multi-site cohort
The BrainGate consortium spans Brown, Stanford, Massachusetts General Hospital, and Case Western in addition to UC Davis, so this is genuinely multi-institution science, not one lab’s pet project
Neuralink bets on channel count and automation
Neuralink’s approach starts from the opposite instinct: pack in as many electrodes as physically possible and automate the surgery so it can scale. The company’s N1 implant carries 1,024 electrodes spread across 64 ultra-thin polymer threads, inserted through a small hole drilled in the skull by a purpose-built surgical robot, then sealed under the scalp so nothing protrudes and data moves wirelessly to an external decoder.
Neuralink is currently running two relevant human trials. The PRIME study targets cursor and device control for people with quadriplegia, and it received its FDA investigational device exemption in May 2023. The VOICE trial is narrower and more ambitious: it targets speech restoration specifically, and it carries its own FDA Breakthrough Device Designation. In January 2026, an ALS patient named Kenneth Shock became the trial’s second participant; by April, Neuralink had demonstrated Shock speaking through the implant in a reconstruction of his own pre-diagnosis voice, built from recordings made before he lost the ability to talk. The system decodes cortical activity into phonemes, then assembles those phonemes into words in near real time.
Ever notice how fast these companies like to quote total patient counts as a proxy for progress? It is worth asking what is actually being measured. Neuralink says it had reached 26 total participants across its studies by June 2026, up from 21 in January, with trial sites now open in the UK, the UAE, and Canada alongside the original US program. A few things to keep in view before treating that scale as equivalent to safety or efficacy data:
More participants is not the same claim as more published, peer-reviewed outcome data
The 1,024-electrode density is the highest of the three approaches here, which raises the ceiling on signal resolution but also raises surgical stakes
VOICE trial results reported so far describe individual patient demonstrations, not aggregated accuracy statistics across a cohort
Neither PRIME nor VOICE has moved past the investigational device stage; there is no commercially available Neuralink product
The robotic, automated insertion process is itself part of what Neuralink is testing, not just the electrode array
Synchron’s minimally invasive alternative
Synchron takes a third path entirely, and it is the one that best illustrates why electrode count alone is a poor way to judge these systems. The company’s Stentrode device carries just 16 electrodes, mounted on a self-expanding, stent-like scaffold. Instead of cutting through the skull, doctors thread it up through the jugular vein into the superior sagittal sinus, a major blood vessel that runs along the top of the brain, in a roughly two-hour procedure that looks far more like a cardiac catheterization than neurosurgery. No craniotomy, no exposed brain tissue, no robotic drill.
That tradeoff bought Synchron an FDA Breakthrough Device Designation back in August 2020, and the company has since run two separate trials: the SWITCH trial in Australia, in collaboration with the University of Melbourne, and the COMMAND trial in the US at Mount Sinai Hospital, alongside Carnegie Mellon and the University of Pittsburgh Medical Center. SWITCH enrolled four patients with ALS or primary lateral sclerosis and published twelve months of safety data with no serious adverse events, along with a two-class signal decoding accuracy of 85.2 percent. Patients used that signal, unsupervised, at home, to text, send email, and shop online, which is communication in the practical sense even without synthesized speech.
Would you trade electrode density for a dramatically safer implantation procedure if the functional result still let you text your family? That is the actual bet Synchron is making, and the accuracy numbers suggest it is not a foolish one. Some specifics worth knowing:
Synchron’s COMMAND trial is a US early feasibility study, planned for six patients, and remains in progress
The company raised a $200 million Series D in November 2025, pushing its valuation toward $1 billion, a sign investors think the endovascular approach can reach commercial approval
Stentrode’s 16 channels support device control and typing, not the phoneme-level speech synthesis Neuralink and BrainGate2 are pursuing
SWITCH trial results were published in the Journal of NeuroInterventional Surgery, giving Synchron a peer-reviewed safety record ahead of the other two
Fewer electrodes means a fundamentally lower-resolution read of neural activity, a real limitation if the goal is eventually full natural speech rather than device control
What communicate means today, and what it will take to get further
Here is the sentence that belongs at the center of any honest article on this topic: none of these three systems is FDA-cleared, and none is commercially available. Every patient using one is enrolled in an active clinical trial under an investigational device exemption. If you are building or investing in this space, that is the fact to anchor every other claim against, because press releases about “communication breakthroughs” and regulatory reality are currently two different timelines running in parallel.
The comparison across these three programs also undercuts a lazy assumption worth retiring: that more electrodes automatically means a better outcome. Synchron gets meaningful device control out of 16 channels delivered without touching the skull. Neuralink is stress-testing whether 1,024 channels and full surgical automation can eventually decode natural speech at scale. BrainGate2’s academic consortium is proving that even a mature, well-validated 256-channel system still needs real-world, in-home testing before anyone should call it ready. Each approach is optimizing for a different variable, accuracy, invasiveness, or scalability, and it will likely take results from all three before regulators have enough data to clear anything for the market.
So which tradeoff would you want if it were your voice on the line: the highest possible resolution, the lowest possible surgical risk, or the longest published safety track record? That is the question this field will actually have to answer before any of this reaches a pharmacy shelf.