An artificial intelligence model that analyzes about 20 seconds of a person reading aloud identified most people with type 2 diabetes in a large UK study. It also raised a false alarm for nearly half of the people whose blood tests showed they did not have the disease. The findings are being presented at the annual meeting of the European Association for the Study of Diabetes (EASD) in Milan, Italy, which runs through Oct. 2.
The results are early. They come from a conference presentation, not a peer-reviewed journal article, and nothing in the researchers' announcement indicates the tool has been cleared by the U.S. Food and Drug Administration. It is not available as a clinical test in the United States.
Still, the research targets a real gap. According to the CDC's National Diabetes Statistics Report, updated Sept. 16, 2026, about 11.0 million U.S. adults with diabetes, or 27.6%, do not know they have it. Another 115.2 million adults have prediabetes. A screen that works over a phone could reach people who never get a blood test, but only if it does not send large numbers of people without diabetes for unnecessary follow-up tests.
Inside the 20-Second Voice Screen
Researchers at the London-based technology company Thymia, working with RMIT University in Melbourne, Australia, trained the model on 63,283 voice samples from 21,129 people in the UK and the U.S. Earlier research has linked type 2 diabetes to voice changes such as increased hoarseness and roughness and weaker breath control while speaking.
To test the model, participants made remote 20-second recordings while reading one of Aesop's fables. The first evaluation included 7,319 UK adults, 217 of whom said they had type 2 diabetes. The model gave a higher risk score to people who reported having type 2 diabetes than to those who did not 80% of the time, a level the researchers considered clinically useful.
The second evaluation is more telling because it checked voice results against blood tests. A subgroup of 801 participants took an at-home HbA1c test within three months of recording. HbA1c reflects average blood sugar over the previous two to three months and is one of the standard tests used to diagnose type 2 diabetes.
Accuracy and False Alarms, Side by Side
Measured against blood test results, the model's accuracy was 75%. Its sensitivity was 82%, meaning it caught about 82 of every 100 people who actually had type 2 diabetes. Its false-positive rate was 47%, meaning about 47 of every 100 people without diabetes were still flagged as possibly having it.
Put simply, the tool missed about 18% of true cases while wrongly alerting nearly half of people who did not have the disease. Every screening tool produces some false alarms, because its job is to decide who needs a confirming blood test. A 47% rate, however, means many people would receive a worrying result and need follow-up testing only to learn they do not have diabetes.
There was a reassuring signal at the other end. None of the people the model placed in its low-risk group had blood results in the diabetic or prediabetic range. The researchers suggest doctors could eventually use short recordings to triage patients and send those at higher risk for blood tests.
Giedrė Čepukaitytė, a research scientist at Thymia, stressed that the tool is meant to sit alongside standard testing. "It is not a replacement for a blood test, and it should never stop anyone who thinks they need one from getting one," she said. She added that the team's next step is to test the model in clinical settings and to understand how well it works for every group of people.
Gaps in the Evidence and Who Could Be Missed
The model performed well across sexes and age groups, but it was less accurate on recordings from Black participants. The authors said that was likely because few Black participants reported having type 2 diabetes. That gap matters in the United States, where the CDC lists race and ethnicity among diabetes risk factors, with African American, Hispanic or Latino, American Indian, and Alaska Native people at higher risk.
Accuracy also dropped for people with heart disease, high blood pressure, or obesity. The researchers said these conditions often occur alongside type 2 diabetes and may cause similar voice changes. That is a concern because people with these conditions are among those most likely to need screening.
Other limits are worth noting. Only 801 of the 7,319 people in the main test had blood confirmation, and the first analysis relied on self-reported diagnoses, which cannot capture people with undiagnosed diabetes. The evaluations involved UK adults, so it is unclear how the model would perform across U.S. accents, languages, and clinics. The study also did not show that voice screening improves health outcomes, and independent replication has not been reported.
Screening Steps That Work Today
People worried about diabetes do not need to wait for new technology. The CDC advises people with risk factors to ask their doctor about getting tested for diabetes. Those risk factors include being 45 or older, having overweight or obesity, being physically inactive, and having a family history of type 2 diabetes. The U.S. Preventive Services Task Force (USPSTF) also recommends screening adults 35 to 70 with overweight or obesity. The HbA1c and fasting blood sugar tests are widely available.
Type 2 diabetes often causes no symptoms at first, which is why testing matters. Common diabetes symptoms listed by the CDC include frequent urination, increased thirst, blurry vision, unexplained weight loss, fatigue, and cuts or sores that heal slowly. Anyone with these signs should contact a clinician. Very high blood sugar with vomiting, fruity-smelling breath, or trouble breathing can signal diabetic ketoacidosis, a condition that needs emergency care.
Many insurance plans cover diabetes screening for at-risk adults as preventive care, and community health centers offer sliding-scale fees for people without coverage. Apps or websites that claim to detect diabetes from your voice deserve caution, since this approach has not been validated in U.S. clinical practice.
Key Questions Answered
What did the voice study find?
An AI model analyzing about 20 seconds of speech caught 82% of people with type 2 diabetes in a blood-tested subgroup of 801 people, but it also flagged 47% of people who did not have the disease.
Can I use this voice test now?
No. The findings were presented at a medical conference, have not been published in a peer-reviewed journal, and there is no indication the tool has FDA clearance.
Who might the tool miss or misjudge?
Accuracy was lower for Black participants and for people with heart disease, high blood pressure, or obesity, according to the researchers.
How is type 2 diabetes diagnosed today?
Through blood tests such as HbA1c or fasting blood sugar, ordered by a health care provider.
Who should ask about screening?
The CDC urges anyone with risk factors, such as being 45 or older, being overweight, being inactive, or having a family history of diabetes, to ask a doctor about testing. The USPSTF recommends screening adults 35 to 70 who are overweight or obese.
What symptoms should prompt a doctor visit?
Frequent urination, increased thirst, blurry vision, unexplained weight loss, fatigue, or slow-healing cuts or sores.
Published by Medicaldaily.com
