A Voice Without a Scalpel: My Bet on the Other BCI Frontier

A person wants to say, “Wait. I have something to add.”

That sentence is small enough to disappear inside a technology demonstration. It is large enough to return someone to a family conversation. Brain–computer interfaces should be judged partly by that scale: the human interruption, the joke delivered in time, the decision expressed without another person guessing.

I expect non-invasive interfaces to produce some of the biggest long-term breakthroughs because technologies that avoid surgery can, in principle, reach far more people. That is a forecast about reach and invention. Today’s highest-bandwidth speech demonstrations often depend on implanted sensors. Hope does not erase that technical difference.

Get the number right

A claim circulating in the brief for this series attributed roughly 75.3 words per minute to Neuralink speech restoration. I could not verify that exact number and attribution in a primary source for this article. It should not be repeated as an established result.

There are important verified achievements without it. A 2024 New England Journal of Medicine study reported an intracortical speech neuroprosthesis used by a participant with ALS for self-paced conversation at approximately 32 words per minute, with accuracy sustained near 97.5% after additional training. This was a specific research system in a particular participant, not a generic performance guarantee for all BCIs.

A 2025 Nature Neuroscience study demonstrated streaming brain-to-voice synthesis, addressing the delay between attempted speech and audible output. Its findings included freeform speech-synthesis measurements around 46.5 words per minute under the reported experimental conditions. Another 2025 Nature paper investigated instantaneous voice synthesis from an implanted interface.

These are distinct systems with different tasks and evaluation methods. A speed number without its conditions can mislead as easily as a wrong number.

What the language model contributes

A neural recording is not a sentence waiting to be copied. A decoder estimates what the person attempted to produce from a noisy signal. Language models can improve that estimate by using the regularities of language.

This is powerful and delicate. The model may turn imperfect evidence into a more usable communication channel. It may also prefer a likely sentence over an unusual sentence the user actually intended. A successful interface therefore needs ways to correct, reject, and confirm output.

The right objective is faithful communication. The most grammatical sentence is not necessarily the person’s sentence. A system helping a patient speak must leave room for surprise, disagreement, minority languages, names it has never seen, and words its statistical expectations would not predict.

Non-invasive means several different things

EEG measures electrical activity from the scalp. MEG measures magnetic signals associated with neural activity. fMRI measures blood-oxygenation changes. These approaches have very different hardware, timing, signal quality, and deployment constraints. Avoiding an implant does not automatically make a system portable or affordable.

The 2023 non-invasive semantic decoding study by Tang and colleagues reconstructed aspects of continuous language from fMRI recordings. It was a remarkable demonstration of semantic information in non-invasive measurements. It was not a consumer headset reading arbitrary private thoughts. The training and cooperation requirements, scanner infrastructure, and reconstruction limitations are essential to understanding the result.

Surface electromyography is another promising route for some communication applications, but it records muscle activity rather than directly recording the brain. Calling every silent-speech interface a BCI blurs the engineering problem.

Why I still favour the non-invasive bet

My case is practical. A device that people can put on and remove could reduce barriers associated with surgery and long-term implantation. It could make repeated design iteration easier. It could support applications whose required bandwidth is lower than fluent speech: selection, confirmation, accessibility controls, rehabilitation feedback, or a few reliable commands.

Those intermediate uses matter. A technology need not transmit a novel every minute to increase someone’s agency. Reliability, comfort, calibration burden, and the ability to stop can outweigh spectacular peak speed.

The hard problems are substantial. Scalp signals are weak and mixed. Motion and muscle activity contaminate measurements. Performance may vary across people and days. A laboratory result with extensive calibration may not survive a crowded home, a clinic, or a matatu journey.

The interface should make a person easier to hear without making them easier to overrule.

Build for African users from the beginning

A Nairobi programme could contribute multilingual evaluation, affordable hardware research, and accessibility design shaped by local users. It should not inherit the assumption that English sentences spoken under ideal conditions represent the whole communication problem.

The research questions are immediate. How long does calibration take? Can users correct the system easily? Does performance degrade with sweat, hair differences, motion, or power interruptions? Can maintenance happen locally? What happens when the internet disappears?

Privacy must also be designed into the product. Neural and physiological data are intimate. Local processing where feasible, explicit consent, clear retention limits, and the ability to delete records should be treated as features that preserve agency.

I am optimistic about AGI helping discover better decoders, materials, and experimental designs. I am more interested in the person finally entering the conversation. The breakthrough arrives when the machine stops being the main character and someone says, unmistakably, what they meant.

Editorial note: This essay expresses the founder’s ideas, developed with AI-assisted research and writing and reviewed against the linked sources on 29 October 2026. Convictions, proposals, and forecasts are identified in the text; scenes framed as imagined are illustrative.

Cover photograph / illustration: Chris Hope; white-balance and noise-reduction adjustments by Colin. CC BY 2.0. Original source. Illustrative EEG recording photograph by Chris Hope, edited by Colin (white balance and noise reduction), CC BY 2.0. Not a participant or device from the studies discussed.