Our voice probably contains more information about our physical and brain health than we think. Researchers have recently built an AI-based tool that estimates people’s age by listening to how they speak; an estimate that may provide clues about the rate of ageing and even the risk of dementia.
The tool does not only pay attention to tone or pitch, but measures hundreds of features, from speaking rate and pauses to word choice and ways of expressing emotions. Researchers, testing this model on nearly three thousand Spanish-speakers, found that the greater the AI-estimated age compared with a person’s real age, the higher the likelihood of cognitive impairment, dementia and accelerated biological ageing. These people had also experienced greater social and economic deprivation. The results of this study were published on Wednesday, 30 September in the journal Science Advances.
According to Live Science, the basic idea of this research is similar to “ageing clocks”; methods that try to estimate a person’s biological age and the probability of age-related diseases by examining bodily biomarkers. The difference here is that the new tool does not require blood tests or brain imaging for this work and obtains the information it needs from a person’s voice.
Adolfo García, one of the authors of this study and director of the Cognitive Neuroscience Center at Universidad San Andrés in Argentina, says existing methods for measuring ageing are very powerful, but many of them rely on brain imaging or blood tests-tests that can be expensive, invasive, or difficult to repeat regularly. According to him, using speech largely overcomes these limitations.
Of course it is still too early to use this tool to diagnose individuals’ dementia risk. A specialist told Live Science that this method must be tested in other languages and in studies that follow people for a long period.
How does this tool estimate age through speaking?
For this study researchers used data from the RED-LET Consortium (a large project on dementia in Latin America). Information from more than 2,900 Spanish-speakers from Argentina, Chile, Colombia, Mexico and Peru was included. Participants were between 18 and 88 years old. About 1,500 were cognitively healthy and the others had mild cognitive impairment, Alzheimer’s or frontotemporal dementia.
Each participant completed seven tasks. For example, they were asked to describe an animated video, name as many vegetables, animals and different words as possible in 60 seconds, and retell a short story once immediately after hearing it and again 20 to 30 minutes later.
The participants’ voices were recorded while performing these tasks. Researchers then examined more than 700 features in these samples; from pauses and speaking speed to pitch, vocabulary range and ways of expressing emotions.
These data were given to a machine-learning model to learn to guess people’s ages solely from how they speak. Researchers then compared the number produced by the AI with each person’s real age and called the difference between the two the “speech age gap.”
The result was clear. In healthy individuals, the age estimated by AI from speech was usually closer to the real age, but this gap increased in those with mild cognitive impairment or dementia. The largest differences were seen in people with frontotemporal dementia, which primarily disrupts language and communication ability.
This association was not only seen in AI results. Those whose speech appeared “older” than their real age also performed worse on memory, language and attention tests and found daily activities more difficult.
Blood tests painted a similar picture. The greater the gap between a person’s real age and the age estimated from their speech, three epigenetic markers also indicated faster bodily ageing. These markers, by examining chemical changes on DNA, provide a picture of the body’s biological age; an age that does not necessarily equal the number of years since a person’s birth.
In Alzheimer’s patients as well, the larger this age gap, the higher the blood level of p-tau 217; a protein that is one of the markers associated with Alzheimer’s disease.
Researchers also found another meaningful correlation. Factors such as adverse economic conditions, food insecurity, limited access to health services, difficult childhood experiences and low educational level were all more prevalent among people whose voices sounded older than their real age. According to García, despite social and cultural differences across the five countries studied, this pattern was the same and reproducible in all of them.
An older-sounding voice does not necessarily mean dementia
Dr. Manisha Parulkar, director at the Memory Disorders and Brain Health Center at Hackensack University Medical Center in New Jersey, who was not involved in this research, considers this method an important step toward making brain health assessment easier and more accessible.
But she cautions against hasty generalization or exaggerated interpretation of the results. Human voice is strongly affected by physical and mental state. Factors such as severe depression, extreme fatigue, or exposure to heavy life stressors may also make a person’s speech pattern appear “older.” For this reason, Parulkar says that as a physician she would never rely solely on such a tool to make a definitive dementia diagnosis.
She says this technology in the future could play more of an alarm role; if the result is unusual, it would prompt the physician to conduct a more detailed examination of the person’s memory and other cognitive abilities.
The study has another important limitation. Many participants were assessed only once and researchers did not track changes in their voices over several years. Therefore it is not yet possible to say whether an ageing speech pattern truly predicts future cognitive decline or exactly how voice and speech change as dementia progresses.
On the other hand, the AI in this study was trained only on the voices of Spanish-speakers from five Latin American countries. Therefore the same model cannot be applied to other languages without testing and modification.
Parulkar says a model trained in Spanish cannot simply be applied to English samples. Each language has its own characteristics and these differences must be considered when designing and testing a model.
Researchers now want to test this method in other languages and follow people for longer periods. The goal is to determine whether changes in how people speak can serve as an early warning for future cognitive decline before symptoms appear.
García says he believes there are strong signals in how people speak that one day might be used to predict future dementia risk, but more and more diverse studies are still needed to reach that conclusion.