How artificial intelligence can tell if you’ve got type 2 diabetes just by listening to your voice – and what this breakthrough could mean for medicine

A 25-second recording of someone’s voice can be used to help identify if they have type 2 diabetes, a new study suggests.

This is based on a breakthrough discovery that people with this common condition have a distinct ‘voice signature’ – this, say researchers at the Luxembourg Institute of Health, is due to the effect the condition can have on the lungs, muscles and nerves involved in producing vocal sounds.

Now they’ve developed an artificial intelligence (AI) algorithm that can pick up the vocal signs such as subtle hoarseness.

The same team are also researching how changes in the voice can help identify some mental health issues such as depression, while other groups of researchers are looking at vocal changes as a way to detect and monitor Parkinson’s disease.

Around 1.2 million people in the UK have type 2 diabetes without knowing, according to the charity Diabetes UK. Left untreated, it can lead to complications such as heart disease and nerve damage. However, while type 2 diabetes can be diagnosed with tests to measure blood sugar, symptoms – such as needing the loo more often – can be subtle, and some show no symptoms, so don’t think to get tested.

The researchers in Luxembourg say their new voice test could be used by people at home to record their voice and check for signs of the disease, which could then be confirmed by a blood test.

It looks for ‘vocal biomarkers’ that the institute’s research has shown are unique to those with type 2 diabetes, says Dr Guy Fagherazzi, the director of precision health.

As he explains, type 2 diabetes reduces pulmonary function (how well your lungs are working) and to produce speech you need to take in enough air to travel through the voice box, where it causes vibrations. The process of turning these into speech involves small muscles in the neck and throat that are controlled by nerves which may be damaged by type 2.

As a result, type 2 can cause changes to the voice such as ‘hoarseness and strain’, and differences in vowel sounds such as ‘ah’, says Dr Fagherazzi.

The researchers ‘trained’ AI to identify vocal signs of type 2 from voice recordings. This was tested using recordings from 607 people reading aloud for less than 30 seconds (taken from a database, Colive Voice, which stores voices for research).

The algorithm correctly identified 71 per cent of men and 66 per cent of women with confirmed type 2 diabetes, reported the journal PLOS Digital Health last month.

The team are also working on biomarkers for other conditions – mental health is an important domain where voice ‘may be key’, says Dr Fagherazzi. ‘Depression can affect many voice characteristics which, when combined, create a different voice. This can be a monotone, with less energy and a slower rate of speech.’

Last year, researchers at the University of South Australia and the Middle Technical University in Baghdad in Iraq published a paper suggesting vocal changes are an early indicator of Parkinson’s, with variations in pitch, articulation and rhythm due to weaker control over the vocal muscles.

They said AI models can detect these changes long before visible symptoms of Parkinson’s appear – in one study, AI was 99 per cent accurate in identifying people with Parkinson’s from their voice.

But Dr Marc Atkin, clinical lead of the diabetes and endocrinology at the Royal United Hospital in Bath, says: ‘The numbers used in this trial are very small, only around 600 people. And what benefit does this have over asking someone to fill in a short questionnaire about risk factors, for instance.’