Researchers at deep-tech company thymia — led by senior machine learning researcher Roseline Polle and senior research scientist Dr. Elisa Brann — built the tool with RMIT University in Melbourne, Australia. They trained it on 63,283 voice samples from 21,129 people in the U.K. and U.S. who had disclosed their diagnosis. They then had participants record themselves reading one of Aesop's fables aloud for about 20 seconds.

How well did it work for different groups?

The model performed consistently across sexes and ages. But it underperformed on recordings from Black participants, likely because too few Black volunteers reported having type 2 diabetes in the training set.

Accuracy also dropped for people with heart disease, high blood pressure, or obesity — conditions that frequently overlap with type 2 diabetes and may produce similar vocal changes.

What would this change for screening?

Current screening relies on blood tests or visits to a primary care doctor. In the U.K., the NHS offers diabetes checks every five years to adults aged 40 and up, but only 40.4% of eligible adults attend. The NHS spends about £10.7 billion on diabetes each year, with around 60% going toward managing complications.

Giedrė Čepukaitytė, a research scientist at thymia who will present the findings at EASD, said the model could reach people who never make it to a clinic.

"This is the largest real-world study of speech-based screening for type 2 diabetes to date that also checks the model's predictions against blood test results as well as against what people reported about their own diagnosis," Čepukaitytė said. "Those flagged as higher risk by the model had blood results to match."

She added: "A speech sample can be taken over the phone or through an app, so we can reach far more of the people who need a blood test than current approaches do, particularly those who never get to a health check."

What are the limitations?

• The model is a possible triage step, not a replacement for blood testing.
• The 47% false-positive rate means nearly half of flagged-but-healthy people would still need confirmatory blood work.
• Performance was uneven across racial and health subgroups.
• Findings have not yet been peer-reviewed; the work was accepted for conference presentation only.

What's next?

The team plans to test the model in clinical settings and to evaluate how well it works for every demographic group before any rollout.

"Our model opens a new route to screening for diabetes," Čepukaitytė said. "It is not a replacement for a blood test, and it should never stop anyone who thinks they need one from getting one. Our next step is to test the model in clinical settings and to understand how well it works for every group of people, because a screening tool has to work for everyone."