Smartphone Voice Clips Track Type 2 Diabetes, With an Age Catch

For two weeks, up to six times a day, adults in India opened an app on their own phones and read aloud whatever random sentence it served them. A transformer model trained on 14,759 smartphone voice recordings from 461 adults told people with type 2 diabetes apart from people without it with a balanced accuracy of about 72 percent in men and about 71 percent in women (guessing scores 50). That is the headline number, anyway. The details are messier.

Balanced accuracy averages two things: how often the model catches people who have diabetes, and how often it correctly clears people who do not, so a lopsided guesser cannot flatter itself. The same Klick Labs group reported higher raw accuracy in 2023 (89 percent in women, 86 percent in men), but the company itself says those figures should not be compared with the new ones, because the yardsticks, inputs and validation differ.

Reading Random Sentences Into a Phone

Earlier voice work leaned on hand-picked measurements such as jitter (tiny wobbles in pitch) and shimmer (tiny wobbles in loudness). This one converts each clip into a spectrogram, a picture of which frequencies are loud at which moment, and hands that image to an Audio Spectrogram Transformer first taught on a big library of everyday sounds, not speech. Because the model reads the whole picture, it can pick up combinations across frequency and time that nobody thought to list in advance. Men and women were modeled separately, and the data were split by person, so none of one person’s recordings could sit on both sides of the training line (a quiet way for voice models to cheat).

Does the picture add anything over the old-style measurements? It did. A baseline built on standard voice-analysis features reached a balanced accuracy of 0.641 in men and 0.628 in women, against 0.724 and 0.713 for the transformer. Independent work lands in similar territory: a Luxembourg-led team using a different voice model on 607 US adults reported broadly similar discrimination, with area under the curve scores (0.5 is a coin flip, 1 is perfect) of 0.75 in men and 0.71 in women.

The team set its threshold to favor catching cases over avoiding false alarms, and the trade-off shows in the women’s results: 0.783 of women with diabetes were flagged, but only 0.660 of those without it were cleared. In practice that would mean extra blood tests for a good many people who turn out to be fine. Fair enough, perhaps, for a screening idea.

The Age Gap Behind the Accuracy

“Now, we can literally see many more clues related to diabetes within vocal biomarkers,” says Yan Fossat at Klick Applied Sciences, the company behind the work and principal investigator of the study. He says the spectrogram approach found combinations of acoustic patterns that “weren’t on anyone’s radar”.

One thing that was on everybody’s radar is age, and it is the awkward part. Men with type 2 diabetes averaged 50.29 years against 34.13 for the controls; for women it was 48.00 against 34.33, a gap of roughly 15 years in each. Voices change with age, and so does diabetes risk, so a model could be listening to birthdays rather than blood sugar. The authors tested that by matching ages strictly: with 69 men in each group, balanced accuracy fell to 0.597, and with 41 women in each group it fell to 0.561. Those matched groups were small, which costs a model training power on its own, yet the paper reads the drop as consistent with age contributing to the full-cohort signal.

Matching still left a residual gap of about a year, and body mass index, higher in the women with diabetes, remains an open confounder, the authors say, and body composition can muddy consumer gadgets too, as when fitness-watch calorie errors grew with body fat in one lab test.

Then there is who was recorded. Everyone came from four clinical sites in India, spoke English, read from a prompt and had no chronic condition other than diabetes, and no outside dataset was available to check the model against (a transformer that read old sleep studies did get a second-cohort check), so the authors expect that performance on other languages, phones and populations could be lower.

A Triage Idea, Not a Test

Klick’s release leads with the 72 and 71 percent figures and does not mention the age-matched rerun, the single-country sample or the missing outside check. The paper, accepted at the Journal of Biomedical and Health Informatics (published by the Institute of Electrical and Electronics Engineers (IEEE)) though not yet in its final edited form, is franker: it calls the work a technical validation rather than a clinical one and says the system should not be used to rule out diabetes or to stand in for guideline-based screening. If it survives prospective testing, the authors see it as a triage signal that nudges someone toward a standard blood sugar test. Whether a phone can ever do that job probably depends on whether the pattern holds up in people who are older, younger, differently accented and recording on other handsets.

For now, the sentence a phone hears is a long way from a diagnosis. Fossat says the group will keep exploring the associations between vocal biomarkers and type 2 diabetes, and a series of studies with Mayo Clinic in Florida, announced in June 2026, is where some of that work is headed.

FAQ