Brain-computer interfaces: reading brain signals to help people speak and control devices

A brain-computer interface, or BCI, reads electrical activity from the brain and turns it into text, speech or commands for a computer. For people who are paralysed or have lost the ability to speak, the aim is a new route to communication, or a way to control a screen or a device without using their muscles. All of the work described here uses implants. The studies are small, and they describe specific people under research conditions, not a device that is available to everyone.

How it works

A BCI records signals, decodes them with a computer model, and sends the result to a screen, a cursor or a voice synthesiser. The model, called a decoder, is trained on a person's own brain activity while they attempt a task, such as trying to say a word.

Recording methods differ. Intracortical arrays use fine electrodes placed inside the brain's outer layer to pick up the activity of nearby neurons. Electrocorticography (ECoG) uses a thin strip of sensors resting on the surface of the brain, beneath the skull. Both need surgery.

The main example: decoding speech

Speech decoding is the clearest worked example so far, so it is worth looking at the studies in detail.

In 2019 Stavisky and colleagues recorded from the motor cortex in two people with tetraplegia after spinal cord injury, both of whom could still speak. The recordings came from a region called the hand knob, which the authors describe as not previously linked to speech. Neurons there were active while the participants spoke and while they moved their tongue, lips and jaw without speaking. Spoken words and syllables could be decoded from single trials. Single-trial accuracy for syllables was 84.6% for one participant and 54.7% for the other, with word accuracy of 83.5% and 61.5% respectively. Chance level in these tests was between 8.6% and 10.1%.

In 2023 Willett and colleagues reported a speech neuroprosthesis for a participant with ALS, who could no longer speak intelligibly. Using intracortical arrays, attempted speech was decoded into text at about 62 words per minute. The word error rate was 9.1% on a 50-word vocabulary and 23.8% on a 125,000-word vocabulary. These figures come from one person, so they show what the system did for that participant, not what it would do for others. Accuracy fell as the vocabulary grew.

A 2026 study in Nature Neuroscience by Brosler and colleagues, conducted at the University of California, San Francisco, recorded from a single high-density ECoG implant in each of three people with paralysis. According to the paper's abstract, participants used parallel speech and gesture decoders to control a personalised virtual avatar, attempting speech and gestures at the same time or separately, and training the decoders on both separate and simultaneous attempts improved performance. A US National Institutes of Health release adds that signals from two of the three participants were used to control a full-body avatar, and that decoders handled mixed speech and gesture better when trained on simultaneous attempts. Neither the abstract nor the release gives accuracy figures. The system was wired to external processing units, and the lead researcher, Edward Chang, calls it a proof of concept in the release. The release also notes, citing earlier research, that attempting speech and gesture at the same time seems to reduce speech ability.

Reaching the brain through blood vessels

Some teams want to avoid open-brain surgery. Synchron's Stentrode is delivered through the jugular vein and placed in the superior sagittal sinus, a large vein beside the brain's movement area. The SWITCH study, published in JAMA Neurology in 2023, enrolled five patients at the Royal Melbourne Hospital in Australia. One was withdrawn before implantation because of a finding on a pre-implant MRI scan, so four received the device and were analysed over 12 months. No device-related serious adverse events leading to death or permanent disability were reported. Eight mild device-related effects occurred across the four patients, including bruising at incision sites, and all resolved without treatment. Each patient used the system to control a computer. Two patients died from ALS complications after the study ended, which the authors judged unrelated to the device. The study was funded by Synchron, which also took part in its design and analysis.

Synchron also announced in July 2022 the first US implant of its device, in the COMMAND trial at Mount Sinai West in New York. That is a company press release, and the claims in it, including the description of the implant as a first, have not been checked independently.

Three states of evidence

Demonstrated, in small studies. Speech decoded from one participant with ALS at about 62 words per minute. Words and syllables decoded from single trials in two people with tetraplegia. Safety results over 12 months for four implanted patients with the Stentrode.

Reported, but not independently checked. Synchron's account of its first US implant. The 2026 avatar result and the decoder training finding, which we have read only in the paper's abstract and the NIH release, not in the full paper.

Speculative. Restoring everyday conversation for people who have lost speech after ALS or stroke. Controlling computers, wheelchairs or household devices for people with severe paralysis outside research settings. Use in conditions beyond paralysis, which none of the studies here tested.

Limits and open problems

• Very few people. The speech results come from a handful of participants, all with severe paralysis or ALS. Results from one or two people do not show how the method would work across a wider group.
• Accuracy. In the 2023 study, decoding accuracy fell as vocabulary size grew.
• Surgery. Intracortical and ECoG systems need surgery that opens the skull, with the usual surgical risks. The blood-vessel approach avoids that, but its published safety data covers four implanted patients, which is too few to rule out rare harms.
• Durability. The SWITCH safety data cover 12 months. These sources do not show whether implants keep working safely for many years.
• Wires and power. The 2026 system was wired to external equipment. The release says a wireless, fully implantable version will be tested, but it has not been shown to work.
• Cost and access. The sources do not address what these systems would cost or who would pay for them.
• Privacy of neural data. Brain signals are highly personal. None of the sources says how such data should be stored, shared or protected, and that needs answering before wide use.

What to watch next

• The full 2026 Nature Neuroscience paper, with its accuracy figures and participant details.
• Independent reporting on Synchron's COMMAND trial and any longer follow-up from the SWITCH study.
• The first tests of the planned wireless, fully implantable system.
• Trials with larger groups of participants, across more than one research site.

This article is general information, not medical advice. Anyone considering a clinical study should speak to their own clinician.