Your Voice Could Flag Diabetes Risk, Study Finds

Your next diabetes screening might not involve a needle at all. Voice Could Flag Diabetes Risk from nothing more than a few short sentences spoken into a phone, according to new research from scientists at Klick Labs whose latest study appeared in the IEEE Journal of Biomedical and Health Informatics this month.

The team converted nearly fifteen thousand smartphone voice recordings from four hundred sixty-one adults into visual maps of sound frequencies, known as spectrograms, and trained an audio transformer model to spot patterns linked to the condition. The only inputs were the recordings themselves and each participant's sex — no age, no body-mass index, no fixed script to read aloud.

On voice alone, the sex-specific models reached balanced accuracy of about seventy-one percent for women and just over seventy-two percent for men, with discrimination scores landing near seventy-nine percent and seventy-eight percent respectively. The researchers stress that balanced accuracy gives equal weight to correctly identifying people with and without the condition, so the figures cannot be compared directly with conventional accuracy scores from earlier work.

The study builds on a 2023 predecessor that asked people to read one fixed sentence while also feeding the model their age, body-mass index and other basic health information. That earlier project surfaced fourteen predefined acoustic features — pitch, intensity, shimmer and jitter among them — tied to the condition, and reported conventional accuracy near eighty-nine percent for women and eighty-six percent for men. The new paper explicitly warns against stacking the two sets of numbers against each other, because the inputs, methods, validation approaches and datasets all differ.

"Now, we can literally see many more clues related to diabetes within vocal biomarkers," said Yan Fossat, the study's principal investigator and the research group's senior vice president of digital health research and development. By capturing relationships across frequencies and time, he said, the analysis moved beyond shimmer, jitter and other predefined voice features and found complex combinations of acoustic patterns that had escaped earlier notice. The full release is available through the company's announcement.

Why screening by voice matters

For readers, the appeal is practical. Finger-prick tests and lab visits are cheap in theory but easy to postpone, and early type 2 diabetes often announces itself with no symptoms at all. A phone-based check that slots into an ordinary day could push detection earlier, when lifestyle changes and treatment still carry the most leverage. Because the design deliberately avoided fixed voice samples, future tools would not need people to read the same sentence under controlled conditions — randomly selected short English sentences were enough, according to the published findings.

The Voice Could Flag Diabetes Risk approach also points to a broader shift in how screening tools are built. Rather than hand-picking the acoustic features to measure, the team let the model learn directly from the spectrograms, which is how it found patterns nobody had thought to look for. It is a small but telling example of how diagnostic research is moving from expert-defined checklists to patterns discovered by the data itself — a trend already visible in everything from a blood test that flags hidden liver damage to routine blood counts that may reveal hidden radiation exposure.

The caveats are real

The usual cautions apply with force. Accuracy in the low seventies means a meaningful share of people would be misclassified, so any voice-based tool would sit at the front door of screening, not replace blood work. The study also leaned on English sentences recorded on smartphones, which leaves open questions about other languages, noisier real-world settings, and older or cheaper microphones.

Voice data itself raises a privacy question that step counts never did. A recording that reveals disease risk is a different category of personal information, and any consumer version of this technology would need to answer where recordings live, who can access them, and whether users can delete them. Those answers do not exist yet.

Still, the research program is widening. In June 2026 the group announced a series of studies with Mayo Clinic in Florida exploring vocal biomarkers for several chronic diseases and serious conditions, and earlier work has already pushed into hypertension and ovulation, plus the effects of audio compression and glucose swings on the voice. The trajectory suggests the same spectrogram approach could become a general early-warning layer for conditions that subtly change how we sound — long before we feel anything wrong. Voice Could Flag Diabetes Risk is still research, not a product, but the direction of travel is unmistakable.