An AI listened to 3,000 voices and guessed who was aging faster than their years

A machine-learning model trained on recordings from nearly 3,000 Spanish-speaking adults can estimate a person’s age from speech alone, and the people it judged older than their birth certificates were more likely to have cognitive problems, dementia and signs of accelerated biological aging. The group counted 2,928 participants in all, drawn from Argentina, Chile, Colombia, Mexico and Peru, and published the results in Science Advances on September 30.

Agustin Ibanez, professor of brain health at Trinity College Dublin and its Global Brain Health Institute, is the senior author. The voice “appears to contain much more information about aging” than researchers previously recognised, Ibanez said, because it captures both the passage of chronological time and signals from cognition, the brain, systemic biology and even a person’s accumulated social environment.

Seven tasks and more than 700 speech features

The volunteers, aged 18 to 88, completed seven language tasks, including describing videos, naming words and telling stories, according to Live Science’s report on the study. From those recordings the team extracted more than 700 distinct features, among them pauses, speaking speed, pitch, vocabulary and emotional expression, and fed them to a model trained to predict chronological age. About 1,500 of the volunteers were cognitively healthy; the rest included people with mild cognitive impairment, Alzheimer’s disease and variants of frontotemporal dementia.

The design borrows the logic of the epigenetic clocks that the study later uses as a yardstick. A model learns what speech typically sounds like at each age, and then the interesting cases are the people it misjudges in a consistent direction. Someone whose recordings sound a decade older than their birth year is flagged as having a positive gap, and the researchers then asked whether that gap lined up with independent measures of brain and body aging, rather than treating the age estimate itself as the product.

The key output is the speech-age gap, the distance between the age the model hears and the age on record. Per the News-Medical summary of the paper by Hernandez and colleagues, healthy volunteers showed the smallest gaps, while larger gaps appeared in Alzheimer’s disease and the frontotemporal dementias.

Brain scans, blood markers and DNA methylation clocks

What gives the gap weight is what it tracked. According to Neuroscience News, larger gaps went with brain atrophy on MRI, higher plasma p-tau217 and faster aging on three independent DNA-methylation clocks, as well as weaker executive function and memory. Live Science adds that people with wider gaps scored worse on memory and attention tests and showed more biological aging markers in blood.

P-tau217 is a form of the tau protein that has become a leading blood marker for Alzheimer’s. In a Swedish study of more than 1,200 older adults summarised by the National Institutes of Health, a blood test built on p-tau217 identified the disease with 88 to 92 percent accuracy. That a speech model correlates with the same marker is the strongest reason the authors treat it as more than a novelty, though a correlation in one dataset is not a diagnosis.

Adolfo Garcia, who directs the Cognitive Neuroscience Center at Universidad de San Andres in Argentina and is a co-author, is described by Live Science as a co-author of the study. San Andres is a partner in ReD-Lat, the multi-country dementia consortium that pools genomic, neuroimaging and behavioral data from Argentina, Brazil, Chile, Colombia, Mexico and Peru alongside U.S. institutions. A speech test needs only a microphone, which is the practical appeal for regions where MRI scanners and blood assays are scarce.

Cross-sectional data and the Spanish-only limit

The authors describe the tool as an investigational research framework rather than an off-the-shelf diagnostic test. The data are cross-sectional, a single snapshot per person, so the study cannot show that a wide gap predicts who will decline later. That would require following volunteers over time, and the team says longitudinal validation is needed before any clinical use.

The model was also trained and tested only on Spanish speakers. Whether the same pauses, pitch patterns and vocabulary measures behave the same way in English, Mandarin or Swahili is untested, and the authors call for validation across languages, cultures and everyday, unscripted speech.

Outside voices urged similar care. Dr. Manisha Parulekar of Hackensack University Medical Center told Live Science that speech could sound “older” simply because a person is severely depressed, exhausted or under serious life stress, a confounder that a one-time recording cannot separate from true biological aging.

The Science Advances paper therefore rests on a specific sample and a specific limit: 2,928 participants, five countries, one language, and an association with p-tau217, brain atrophy and methylation age that still has to survive a longitudinal test.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.

More from Morning Overview