Researchers at the University of California, Davis demonstrated long-term home use of a speech neural interface by a patient with amyotrophic lateral sclerosis (ALS). Over 19 months, Casey Harrell transmitted 183,060 sentences — 1,960,163 words — at an average speed of 56 words per minute.

According to a paper in Nature Medicine, the authors studied not a laboratory demonstration but the everyday use of the system without researchers present.

What home use revealed

Harrell used the neural interface at home for more than 3,800 hours. The system helped him communicate with family, friends, colleagues and doctors, send messages and emails, take part in video calls, use the internet and maintain full employment despite paralysis.

By the patient's estimate, the system decoded 92% of sentences as at least "mostly correct." In formal tests, where Harrell was shown words on a screen, accuracy exceeded 99% with a vocabulary of 125,000 English words. The peak figure in the paper was 99.2%.

At the time the study was published, Harrell had used the system on 444 of the 653 days after implantation. After his assistants were allowed to connect and disconnect the equipment on their own without researchers present, the average interaction time grew from 3.7 to 9.5 hours per day.

How the system works

In 2023, four microelectrode arrays were implanted in Harrell's left precentral gyrus — the area of the brain associated with speech coordination. The system reads signals from 256 cortical electrodes when the patient attempts to speak.

The algorithm converts neural activity into phoneme probabilities every 80 ms, after which a language model selects the most likely sequence of words from a vocabulary of roughly 125,000 words. The text is displayed on the screen in real time. Once a phrase is finished, the system can voice it with a synthesized voice tuned to the sound of Harrell's voice before his illness.

To control the computer, the patient also used a cursor decoder. It was previously thought that speech and its movements might require different brain regions, but the team showed that both modes can be implemented through signals from the speech motor cortex.

From demonstration to everyday tool

The team had already shown the neural interface's high accuracy in laboratory sessions. In 2024, the researchers reported that in the first 30-minute training session the system reached 99.6% accuracy with a vocabulary of 50 words. On the second day, after another 1.4 hours of setup, accuracy was 90.2% with a vocabulary of 125,000 words.

The new work shifts the focus from controlled trials to long-term everyday use. According to the authors, this is one of the key steps toward practical neural interfaces for people with severe motor impairments. Study co-author Sergey Stavisky stated that the 3,800 hours of brain activity recorded during use of the system became, according to his data, the largest individual dataset with single-neuron resolution.

The authors emphasized that the study describes a single clinical case. It is not yet known to what extent the results can be transferred to other patients, implantation areas, electrode types and neurological conditions. The system also remains experimental. It uses wired connections, requires daily connection by trained assistants and, because of the size of the equipment, is suitable only for home use.

In March, Neuralink chief Elon Musk reported that its BCI device is capable of restoring speech to people who have lost it. One of the patients, by the 80th day of using the neural implant, was able to launch World of Warcraft and play with the power of thought.

In April, it became known about brain-computer device user Galen Buckwalter, who learned to create music with the power of thought.

Recall that in spring the media reported that the startup of a former Neuralink president would test a biohybrid chip on humans.

Source: ForkLog