(Credit: pixs4u on Shutterstock)
Computer Analysis of Speech Links Older-Sounding Voices to Dementia and Brain Aging
In A Nutshell
- A computer model guessed age from speech alone, and its guesses overshot real age by more for people with Alzheimer’s disease or frontotemporal dementia.
- Healthy adults had the smallest speech age gaps, while a language-focused form of frontotemporal dementia had the largest.
- Larger gaps lined up moderately well with brain-scan aging and only weakly with DNA-based aging.
- The study captured one moment in time among Spanish speakers, so it cannot yet say whether speech can predict who will develop dementia.
Voices of people with dementia may sound older than they are, at least to a computer. In a study of nearly 3,000 Spanish speakers, software guessed each person’s age from speech alone, and the guesses overshot real age by more for people with Alzheimer’s disease or frontotemporal dementia, according to research in Science Advances.
Researchers call the tool a speech clock and the difference between its guess and a person’s real age a speech age gap. A positive gap means someone sounds older than their years, while a negative gap means they sound younger. Healthy adults had the smallest gaps. The biggest belonged to people with a form of frontotemporal dementia that mainly damages language.
Those gaps also moved with brain scans, a blood protein tied to Alzheimer’s, and signs of aging in DNA, hinting that speech carries traces of how the brain and body age. But everyone was assessed at a single point in time, with diagnoses already known, so the study cannot say whether a voice could warn of dementia before it starts.
Speech Clock Missed Actual Age by About Nine Years on Average
Participants came from Argentina, Chile, Mexico, Peru, and Colombia, 2,928 in all, recruited through a Latin American dementia research network. About half, 1,504, were cognitively healthy. The rest included 1,068 with Alzheimer’s, 332 with frontotemporal dementia, and 24 with mild cognitive impairment, a stage that often comes before dementia. Frontotemporal dementia is a less common type that can affect behavior, movement, or language. Of those 332 patients, 77 had a form centered on language and 255 had forms that mainly affect behavior or movement. Two-thirds of participants were women, and the average age was 65.
Speech was recorded during seven tasks run by neuropsychologists in quiet rooms. Participants described a short cartoon, named words starting with a given letter, listed animals and vegetables, and retold a short story twice, once right away and once 20 to 30 minutes later.
Talking draws on memory, word retrieval, and mental speed, so researchers expect trouble with any of them to show up in how a person speaks. Software pulled more than 700 measurements from each recording, including speaking pace, pitch, vocabulary, emotional tone, and how much a person said. A computer model then learned to guess age from those patterns. It was a rough clock, missing by about nine years on average. Speaking pace and the amount of speech pushed guesses younger, while pitch and emotional tone pushed them older.
Speech Age Gaps Grew Larger Across Dementia Groups
Gaps in the behavior-or-movement forms of frontotemporal dementia topped those in Alzheimer’s, and mild cognitive impairment also exceeded healthy adults, though only 24 people had it. Results held after accounting for age, sex, education, and country, and when the model was built on four countries and tested on the fifth.
Bigger gaps came with weaker scores on memory, attention, and planning tests. Memory showed the strongest tie, particularly in Alzheimer’s and the behavior-or-movement forms of frontotemporal dementia.
Gaps rose with levels of p-Tau217, a blood protein tied to Alzheimer’s disease, though the connection was modest and showed up only among Alzheimer’s patients when groups were examined separately.
Life circumstances showed a connection too. Participants facing harder social conditions, including less schooling, food insecurity, financial strain, and limited health care access, tended to have larger gaps, a pattern seen in healthy adults and Alzheimer’s patients but not the other groups. The study cannot show that hardship caused the speech differences, only that the two were linked.
Speech Gaps Tracked Brain Scan Aging More Closely Than DNA-Based Aging
Computer models estimated each person’s brain age from two kinds of MRI, one measuring brain structure and one measuring brain activity. People whose speech sounded older also tended to have older-looking brains, a moderately strong link that appeared in healthy adults and every dementia group with both measures.
Links to DNA were weaker. Blood tests that read chemical tags on DNA, which shift as people age, offered another age estimate, and speech gaps lined up with it only slightly, mostly in healthy adults and Alzheimer’s patients. Speech relies on coordinated brain systems, which the authors say may explain why links to brain scans ran stronger than links to whole-body markers.
People who spoke more languages tended to have slightly smaller gaps, but the link was weak and fell short of statistical significance, so a protective effect is only a hypothesis.
Speech Clock for Dementia Screening Still Needs Long-Term Testing
All participants were native Spanish speakers, and speech came from structured tasks, so results may not carry over to other languages or casual chatter. Recording conditions and microphones may have varied across sites, and some participants lacked blood, brain, or social data.
Recording a voice costs far less than a scan or a lab panel, and the authors call speech a “low-cost biomarker candidate” for research in underrepresented settings. If studies that follow people over time confirm the pattern, recordings could eventually help expand screening and monitoring. Until then, a computer can hear that people with dementia, as a group, sound older than their years, but whether it can hear that early enough to matter is still unanswered.
Disclaimer: This article is for informational purposes only and is not medical advice. Anyone with concerns about memory, speech, or brain health should consult a qualified health professional.
Paper Notes
Limitations
Study authors note that the design is primarily cross-sectional, meaning everyone was assessed at a single point in time, which precludes causal inference and limits the ability to track within-person aging or predict future clinical conversion. Differences in recording conditions, microphones, and speech prompts across sites may have introduced residual noise into the speech data, and automated speech-feature tools may embed dialectal or cultural biases that have not been systematically evaluated. Sex, medical comorbidities, and mood-related fluctuations are likely to influence speech production, though sex-adjusted sensitivity analyses showed the main group-level patterns persisted. Sample sizes were limited for some diagnostic groups, especially mild cognitive impairment and language-dominant frontotemporal dementia, and in some biomarker subsets, which reduces statistical power. Biomarker and social exposome data were not available for all participants, raising the possibility of selection bias in the combined analyses. The data are restricted to Latin American cohorts and relatively structured speech tasks, so generalization to other languages, cultures, and more naturalistic conversation needs to be established. Harmonized data on the number of languages spoken were not uniformly available, and the multilingualism result showed a trend only.
Funding and Disclosures
Support came from multiple sources, including the Multi-Partner Consortium to Expand Dementia Research in Latin America (ReDLat), which is supported by the Fogarty International Center and the National Institutes of Health’s National Institute on Aging, the Alzheimer’s Association, the Rainwater Charitable Foundation’s Bluefield Project to Cure Frontotemporal Dementia, and the Global Brain Health Institute. Individual authors also disclosed support from Davos Alzheimer’s Collaborative, Chile’s national research agency (ANID), the Wellcome Trust, Wellcome Leap, and other national and international funders. The paper states that the funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. The authors declare no competing interests, though the acknowledgments list that author Bruce Miller serves on advisory boards for several organizations, receives royalties from academic publishers, and holds editorial roles at two journals, and that author Vaibhav A. Narayan is Chief Data and Digital Officer of the Mental Health Goals Program.
Publication Details
This article is based on the study “Speech clocks decode dementia phenotypes, social exposome, and biological aging,” published in Science Advances on September 30, 2026 (Volume 12, Issue 40, eaef9864). The study was submitted February 2, 2026, and accepted August 25, 2026. Hernan Hernandez and Lizeth Katherine Pedraza contributed equally as lead authors, and Adolfo M. García and Agustin Ibanez are the corresponding authors. The full author list includes researchers from institutions across Latin America, the United States, and Europe. The paper is open access under a Creative Commons Attribution 4.0 license. DOI: 10.1126/sciadv.aef9864.
