A soft, wearable neck device that translates silent throat muscle vibrations into audible speech was associated with improved communication fluency and higher patient satisfaction among stroke survivors with dysarthria in a small clinical study.

The device, Revoice, combines textile sensors with large language models (LLMs) to translate silent articulatory movements into full sentences. Unlike prior silent speech technologies, this wearable supports a continuous, natural flow of communication without pauses that disrupt conversation, a limitation of standard augmentative and alternative communication (AAC) devices.

Stroke patients who used the wearable device achieved low word- and sentence-level error rates, whereas optional artificial intelligence (AI)-assisted sentence expansion significantly improved fluency and emotional expressiveness. Patients also reported a greater sense of ease with communication and higher overall satisfaction compared with the basic word-only output.

“To our knowledge, this is the first wearable assistive communication system that enables real-time (sub-second latency) expressive communication for people with dysarthria,” lead investigator Luigi G. Occhipinti, PhD, Department of Engineering, University of Cambridge, Cambridge, England, told Medscape Medical News.

“Our work demonstrates that subtle, residual muscle and skin movements in the neck region can be harnessed by our wearable intelligent throat [IT] device, as a smart silent speech interface to help restore fluent, naturalistic communication ability in daily life,” he added.

The study was published online on January 19 in Nature Communications.

Communication Gaps Common

Dysarthria, a motor speech disorder, is a common complication of stroke and other neurologic disorders, including Parkinson’s disease and amyotrophic lateral sclerosis. A lack of neuromuscular control over the vocal tract can limit intelligible speech and interfere with social interaction, rehabilitation, and quality of life.

AAC technologies, such as head- or eye-tracking spelling systems and brain-computer interfaces via the use of neuroprosthetics, can help some stroke patients. However, these tools can be slow, cognitively demanding, invasive, or impractical for daily use. Even though some stroke survivors retain partial control of laryngeal or facial muscles, there is a lack of effective tools that can translate those movements into fluent speech.

Wearable silent speech devices have shown promise as a noninvasive alternative to traditional AAC. However, most existing systems rely on fixed time windows or whole-word decoding, which forces users to pause between utterances and limit natural speech flow.

To address these limitations, study investigators used AI-assisted sentence synthesis and expansion to enable continuous, token-level decoding of silent speech and reduce patient burden among stroke patients with dysarthria.

Improved Accuracy and Patient Satisfaction

To evaluate the IT wearable device, they recruited a total of 15 participants, including five stroke patients with dysarthria (mean age, 43 ± 7.8 years; four men and one woman) and 10 healthy control participants (mean age, 25.3 ± 4.1 years; six men and four women).

The investigators used a 47-word vocabulary of common phrases in Chinese and 20 short sentences based on these words. The device simultaneously measured two different signals, including silent speech muscle vibrations and carotid pulse signals (heartbeat). The system utilized 144-millisecond intervals to enable a seamless speech flow.

Aside from being a lightweight IT device, a key factor was the integration of the machine learning-based signal decoding with LLM-driven sentence expansion and emotion status awareness, said Occhipinti.

“Thanks to these advanced functionalities embodied in the device, the system supports natural, low-effort attempts at expression and then intelligently refines and expands them in a personalized and context-aware manner,” he said.

Among the stroke group, the system achieved a word error rate of 4.2% and a sentence error rate of 2.9%. In addition to recognizing speech, the system decoded emotional states, such as frustration or relief from heartbeat signals with 83.2% accuracy. AI-driven sentence expansion was associated with a 55% increase in overall patient satisfaction, with participants reporting that communication felt more natural, shifting from “somewhat satisfied” to “fully satisfied.”

Study limitations included the small sample size, limited tested vocabulary, a short evaluation period, and lack of a larger, more diverse population. These limitations may limit the generalizability of the findings and how the technology would perform across broader clinical settings.

“Our goal is to help patients that are unable to produce complete sentences and communicate fluently, as well as speech and language therapists and caregivers to engage more effectively with the patient,” Occhipinti said.

He added that future work will need to focus on scaling up datasets across more diverse patient populations and improving cross-user generalization.

“With sufficiently large and representative datasets, models are expected to gain stronger out-of-distribution,” Occhipinti added.

‘Promising Research’

Commenting for Medscape Medical News, Jessica Hooke, MA CCC-SLP, CBIS, MPH, American Stroke Association national volunteer expert, clinical rehabilitation specialist for community programs and wellness, Orlando Health Advanced Rehabilitation Institute, said the study highlights the possibilities for communication technology in poststroke care.

“AI-driven devices and AAC have steadily increased and improved over the years due to advancements in technology,” Hooke said. “However, the application of high-tech AAC devices for individuals with dysarthria are often not used during the acute phase and are often introduced after other strategies have been tried.”

“I believe that the research is promising, and hopefully with advancements, this may be a feasible device that could be applied practically in the clinical setting,” she added.

However, Hooke acknowledged several limitations must be addressed before this technology can be widely used in rehabilitation and outpatient care. She noted that the study did not specify the type or level of severity of dysarthria among patients, which could have significantly affected their results.

Time since onset could also influence outcomes, Hooke added. “There was no reference to the time of onset of dysarthria for each participant. For example, did their stroke occur 1-3 months ago or 2-4 years ago?”

She also cited “cultural and language differences” as another potential limitation.

“My hope that as research and resources become more readily available, that they will be affordable, easily accessible, and user-friendly for survivor use,” she said.

The study was funded by the National Natural Science Foundation of China, Beihang Ganwei Project, British Council, the UK Engineering and Physical Sciences Research Council, and Haleon. Hooke reported having no relevant financial relationships.