A path may open for patients whose speech becomes slurred or who struggle to string sentences together after a stroke to communicate naturally without surgery to implant electrodes in the head.
A University of Cambridge team said on the 19th that it developed Revoice, a wearable device in the form of a choker worn around the neck, and restored in real time into sentences what users intended to say even when they could not produce vocal sounds. It makes natural communication possible without an invasive brain-computer interface (BCI) that implants electrodes in the brain.
Dysarthria, which commonly appears after a stroke, is a symptom in which weakened muscles of the face, mouth, and vocal cords blur pronunciation or slow speech. Thoughts are clear but the mouth and throat do not follow, so patients often speak only short, broken words. Many stroke patients have dysarthria, and some also develop aphasia, in which the language-processing function itself is damaged.
Luigi Occhipinti, a professor at the University of Cambridge, said, "Patients know exactly what they want to say, but the signals between the brain and the neck are disrupted by stroke, making physical expression difficult and causing frustration," and "This is a heavy burden not only for patients but also for families and caregivers."
The core of Revoice is ultra-sensitive sensors and artificial intelligence (AI). The necklace-type device captures subtle vibrations of the neck muscles and heart rate signals, and the team designed two AIs to operate based on these. One infers fragments of words that the user mouths silently, and the other gauges emotional states from indicators such as heart rate and adds contextual information such as time of day and weather to complete sentences in real time. A lightweight large language model (LLM) with low power consumption is used for sentence expansion.
For example, if a user only mouths roughly "Our hospital go," the device, based on signs such as elevated heart rate, reflects emotions like frustration or restlessness and the context of a late hour to turn it into a natural sentence such as "It's late, but I still feel unwell. Could we go to the hospital now?"
In a small trial conducted with Chinese researchers on five stroke patients with dysarthria and a control group of 10 people, the word error rate was 4.2% and the sentence error rate was 2.9%. The lower the error rate, the more accurately the device inferred the user's intent. Unlike existing assistive communication devices, which rely on letter input, eye tracking, or invasive brain implants and are slow or burdensome, Revoice produced relatively smooth conversations with just a few words. Participants reported that overall satisfaction rose by 55%.
The team sees the technology as applicable not only to stroke rehabilitation but also to supporting patients with Parkinson's disease and motor neuron disease, and is preparing clinical research on English-speaking patients with dysarthria.
The researchers added, "Before it can be widely used, larger-scale clinical trials and validation in diverse situations are needed," and "We will improve it to support multiple languages, recognize a wider range of emotional states, and operate as a fully stand-alone device for daily use."
The findings were published the same day in Nature Communications.
References
Nature Communications (2026), DOI: https://doi.org/10.1038/s41467-025-68228-9