The State of Speech BCIs

In 2021 a man who had not spoken for more than a decade produced sentences from a fifty-word vocabulary at a median of 15.2 words per minute. It was a landmark, and by the standards of the field today it was slow and narrow.

Two years later, two groups published side by side in Nature. One, recording from intracortical arrays, decoded attempted speech at 62 words per minute over a 125,000-word vocabulary. The other, using a high-density surface array, reached a median of 78 words per minute and added a synthesised voice and an animated face. Roughly one word in four was wrong in both.

The following year a team at UC Davis closed most of that gap. Their participant, a man with ALS, communicated with 97.5% word accuracy and kept using the system for more than eight months and 248 hours.

The intracortical results come from arrays a few millimetres across, placed in the ventral precentral gyrus and recording the spiking of individual neurons. Their strength is information density: a small patch of cortex carries enough articulatory detail for open-vocabulary speech.

The surface-array results come from electrocorticography grids that do not penetrate the brain and cover a much wider area. Their strength is stability and breadth. In September 2026 the same approach was shown to decode speech and upper-body gestures together from one implant.

It would be a mistake to read these as a contest with a winner. They trade resolution against coverage, and the right choice may depend on what else a patient needs the implant to do.

Text is a means, not the goal. In 2025 two papers showed speech synthesised as the participant attempted to speak: one streaming audio in 80-millisecond increments from a surface array, the other producing voice from intracortical activity with enough control over pitch for the participant to ask questions, add emphasis and sing short melodies.

Hearing your own voice as you speak is how speech is regulated. Closing that loop changes a speech BCI from a transcription service into something closer to a voice.

The figures above are not a league table. They come from different participants, different causes of speech loss, different vocabularies and different tasks. A rate measured while copying prompted sentences is not the same as a rate in free conversation; the UC Davis participant's self-paced conversations ran at about 32 words per minute.

The field has responded by publishing data. The recordings behind the 2023 intracortical result anchor a public benchmark on which outside teams have reduced error rates without any new surgery. That is the clearest sign of a maturing research area.

A 2025 study in four participants found that inner speech is represented in motor cortex and can be decoded in real time, and that it shares much of its neural code with attempted speech. The same paper demonstrated strategies that prevent a system from decoding inner speech the user did not intend to share.

This matters for trust. The concern is real, and it has a measurable technical answer that can be specified, tested and regulated.

Until recently every speech BCI result came from an academic study using percutaneous research hardware. That is changing. Paradromics implanted its first Connect-One participant in June 2026 and reported real-time speech in September, without figures. Neuralink's VOICE study, registered in late 2025, lists six planned participants.

Neither has published performance data. When they do, the academic benchmarks provide a ready comparison.

Fully implanted, wireless hardware that matches the accuracy of wired research systems. More participants, across more conditions and languages. Follow-up long enough to show what happens as diseases such as ALS progress. And decoders that work for a new user without hours of calibration.

None of these requires a scientific breakthrough. They require engineering, clinical evidence and time — which is a statement about how far the science has already come.