After a total laryngectomy, the voice doesn’t fade. It vanishes. The vocal folds are gone, and with them the ability to produce even a whisper. For the roughly 175,000 people worldwide who undergo the procedure each year, according to GLOBOCAN data, the options have long been limited: a handheld electrolarynx that produces a robotic monotone, or months of training to master esophageal speech. Neither sounds like the person who used to speak.
A wave of peer-reviewed research published through early 2026 suggests a fundamentally different approach is gaining ground. Soft, flexible sensors worn against the neck can now detect the tiny muscle contractions and skin deformations that occur when someone tries to talk, even silently, and feed those signals to AI models that reconstruct audible words. The technology is still in the lab, but the results so far are striking enough to reshape expectations for assistive communication.
Decoding silent speech through the skin
The clearest demonstration comes from a study published in Cyborg and Bionic Systems, a Science Partner Journal affiliated with AAAS. Researchers built a soft, adhesive patch that wraps around the neck and measures multiaxial strain, capturing the subtle stretching and compression of skin as throat muscles move during attempted speech. An AI decoder trained on those strain patterns then translated them into reconstructed audio with measurable accuracy, all without surgery, implanted electrodes, or any audible input from the wearer.
The concept has roots in earlier work at Pohang University of Science and Technology (POSTECH), where a team laminated ultrathin crystalline-silicon strain gauges onto the face and neck. Their 2022 study in PNAS showed that deep learning could classify distinct words and phrases from surface strain alone, confirming that skin deformation carries enough speech-relevant information to serve as a sensing channel, even when no sound is produced. That foundational finding opened the door for the more advanced reconstruction systems now being tested.
From healthy volunteers to stroke patients
A separate study published in Nature Communications pushed the technology closer to clinical reality. The research team developed a textile-based “intelligent throat” that picks up two types of signal simultaneously: vibrations from the extrinsic laryngeal muscles and pulse patterns from the carotid artery. By training AI on this multimodal data, they reconstructed speech and evaluated it using word error rate (WER) and sentence error rate (SER), standard benchmarks that allow comparison with other speech technologies.
What sets this work apart is the patient population. Rather than testing only on healthy volunteers mouthing words in a quiet room, the team fine-tuned and validated its system on people with dysarthric speech following a stroke. These are individuals with real, variable impairments, not idealized lab subjects. The results offer early but meaningful evidence that neck-worn AI sensors can function where they are needed most.
Reaching people after laryngectomy
For people who have lost the larynx entirely, a different sensing strategy is required. A study in PLOS ONE evaluated silent speech recognition using surface electromyography (EMG) sensors placed on the face and neck of post-laryngectomy participants. Because the vocal folds are absent, the system relies on residual muscle activity associated with attempted articulation to generate text or synthetic speech. The researchers documented vocabulary size, participant demographics, and recognition performance, aiming to provide a communication channel that feels more natural and flexible than an electrolarynx, particularly during conversation.
On the engineering side, a study in Computers in Biology and Medicine showed that a conformer-based deep neural network could synthesize intelligible speech from three-axis accelerometer signals mounted near the throat. The model predicted acoustic features from vibration data, converted spectrograms into audio waveforms, and effectively rebuilt the sound of a voice from non-audio signals alone. Accelerometers differ from strain gauges and EMG, but they share the same core advantage: they capture signals that are far less sensitive to background noise than a standard microphone.
Why the neck beats a microphone in noisy rooms
That noise resilience may be the technology’s most practical selling point. A broad review in Nature Communications surveyed multiple classes of skin-adjacent wearable sensors and found that, under realistic noise conditions, including background chatter, wind, and overlapping voices, throat-mounted devices maintained recognition accuracy where conventional audio-based systems degraded sharply. The authors tested across sensor types and reported on sensitivity, bandwidth, and mechanical flexibility, concluding that the neck has become a focal site for next-generation speech interfaces precisely because it offers a direct mechanical link to articulation that ambient sound cannot drown out.
For someone trying to communicate in a busy hospital ward, a crowded family gathering, or a noisy street, that difference could be the gap between being heard and being ignored.
What still stands between the lab and daily life
None of these systems are ready for the nightstand. Several significant gaps remain, and acknowledging them honestly is essential to understanding where the technology actually stands as of spring 2026.
Long-term wearability is untested. The multiaxial strain study does not report systematic data on comfort over hours of continuous use, the effects of sweat and movement on signal quality, or how adhesive performance and skin irritation change over weeks. Many wearable medical devices that perform well in short trials run into adherence problems when people are asked to wear them all day.
Real-world conditions are largely unexplored. Most published experiments use quiet rooms and scripted prompts, which simplifies the decoding problem considerably. Spontaneous speech, interruptions, shifting posture, and physical activity all introduce variability that current AI models have not been rigorously tested against. How often a sensor would need recalibration after slipping slightly on the skin during a normal day remains an open question.
One device may not fit all causes of voice loss. A person with post-stroke dysarthria retains partial control over laryngeal and articulatory muscles. Someone after a total laryngectomy has lost the vocal folds entirely and relies on different residual musculature. The stroke-focused and laryngectomy-focused studies each optimize for a specific population, but there are few cross-condition comparisons. Whether a universal model can serve these diverse groups, or whether each user will need highly individualized training, is unknown.
Cost and access are unaddressed. None of the current papers provide manufacturing cost projections, expected device lifetimes, or maintenance requirements. They do not outline pathways for insurance coverage or public health system adoption, which are critical for reaching the millions of people globally living with voice loss. Without that information, it is impossible to judge whether these systems will remain research curiosities or become viable alternatives to existing augmentative and alternative communication (AAC) devices.
Patient experience is missing from the data. The studies report objective metrics like WER and SER but rarely include qualitative feedback. How does the reconstructed speech sound to a family member across a dinner table? Is the delay between articulation and output short enough to feel like conversation? Does wearing a visible neck sensor affect social interactions or carry stigma? The emotional weight of hearing a synthetic version of one’s former voice could strongly influence whether someone adopts the technology or abandons it, and that dimension is not yet captured in the literature.
Why neck-worn speech sensors may redefine assistive communication
The peer-reviewed record, taken together, supports a specific and limited conclusion: the skin of the neck carries enough mechanical information about speech to allow AI systems to reconstruct intelligible words, and this works not only in healthy volunteers but in people with real speech impairments. That is a meaningful advance. It is not yet a product, a therapy, or a guarantee.
Readers evaluating these claims should look for primary research with quantitative benchmarks, named patient populations, and transparent descriptions of what was and was not tested. The Cyborg and Bionic Systems paper and the Nature Communications intelligent throat study meet that standard. Promotional summaries that leap from lab accuracy to phrases like “restoring natural conversation” should be treated with skepticism until long-term, real-world trials back them up.
What is clear, even at this early stage, is that the old binary, speak or be silent, is starting to crack. For the people on the silent side of that divide, the neck-worn sensors emerging from these labs represent something that has been scarce in assistive communication for decades: a genuinely new idea, grounded in solid science, with room to grow.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
