A study published in Science Advances has shown that automated analysis of brief voice recordings can yield a “speech age” score that tends to be older in people with dementia than in cognitively healthy adults. The research drew on almost 3,000 native Spanish speakers across five Latin American countries and found that the gap between speech age and actual age correlates with brain-imaging measures, DNA-based aging markers, a blood biomarker for Alzheimer’s disease and lifetime social hardship.
Biological aging clocks aim to estimate a person’s biological age from data such as MRI brain measurements or patterns of DNA methylation. The difference between the predicted age and chronological age — the age gap — is interpreted as accelerated aging when positive. While many clocks rely on costly tests, the study explored whether speech could offer a low-cost, scalable alternative that captures multiple dimensions of aging.
Natural speech may provide a practical, non-invasive window into how the brain and body age, the researchers suggest. Speech requires rapid coordination of cognitive and physical systems, which decline with normal aging and more markedly with neurodegenerative disease. A slower speaking rate, longer pauses and shifts in vocal pitch and emotional tone can signal processing slowdowns and changes in semantic memory.
The project was led by Hernan Hernandez of the Latin American Brain Health Institute at Universidad Adolfo Ibáñez in Chile, with senior authors Agustin Ibáñez and Adolfo M. García. Ibáñez is a professor of global brain health at the Global Brain Health Institute in Dublin and scientific director at the Latin American Brain Health Institute, while García directs the Cognitive Neuroscience Center at Universidad de San Andrés in Argentina and is affiliated with several brain-health institutes.
Researchers sought to test whether discrepancies between chronological age and speech-predicted age could differentiate dementia types and reflect internal biological aging.
The study analysed 2,928 native Spanish speakers from Colombia, Peru, Mexico, Chile and Argentina. The participants had an average age of about 65, with women making up 67 per cent of the cohort. The group included 1,504 cognitively healthy individuals and 1,424 patients with various neurodegenerative conditions, including 24 with mild cognitive impairment and 1,068 with Alzheimer’s disease, plus 332 with frontotemporal dementia, subdivided into language-dominant and behaviour/motor variants.
Each participant completed seven brief speaking tasks in a quiet room, such as describing an animated video, retelling a short story immediately and after a delay, and naming as many words as possible under time pressure. The recordings were analysed with TELL, an automated speech-analysis app in which García serves as founder and chief scientific officer, which removed background noise, isolated the participant’s voice and produced transcripts.
From the audio and transcripts, researchers extracted more than 700 acoustic and language measurements, grouping them into seven composites: speech timing, vocal pitch, emotional content, word frequency, verbosity, semantic concreteness, and semantic granularity (how specific a word is, for example “robin” versus “bird”).
A machine-learning model was trained on these seven categories to predict chronological age, and the speech-age gap was the difference between predicted age and actual age. On average, the model’s age estimates missed participants’ true ages by about nine years and accounted for roughly 44 per cent of the age variation in the sample; after a standard correction for age-prediction, the predicted and actual ages correlated at 0.74. Verbosity and speech timing were the strongest indicators of preserved or younger-sounding speech, while variations in pitch and negative emotional word content were most strongly linked to accelerated speech aging.
When comparing diagnostic groups, cognitively healthy adults had the smallest speech-age gaps, while every patient group showed larger gaps. Language-dominant frontotemporal dementia produced the largest gaps, followed by behavioural or motor variants of frontotemporal dementia, then Alzheimer’s disease and mild cognitive impairment.
Sensitivity analyses that controlled for chronological age, sex, education, country of origin, medical conditions and multilingualism did not overturn the main findings; there was a slight trend toward younger-sounding speech among multilingual speakers, though this did not reach statistical significance.
The full speech clock outperformed any single acoustic or language category in separating the diagnostic groups, suggesting that aging leaves a distributed signature across multiple speech dimensions. “Aging leaves a distributed signature across multiple dimensions of speech—timing, lexical choice, semantic properties, prosody, and verbal output,” García explained. “No single feature tells the whole story; the informative signal emerges from their combination.”
García also cautioned against overestimating speech changes as a simple indicator of health. “The speech clock captured clear differences between groups and showed especially strong links with brain-age measures, but it is best viewed as a signal that becomes informative when combined with other evidence, not as a stand-alone test of a person’s health or prognosis,” he said. “A positive speech-age gap should not be interpreted as a diagnosis or as proof that someone is literally ‘aging faster.’”
Beyond group differences, larger speech age gaps were associated with poorer performance on cognitive tests, particularly those with verbal components, and with higher levels of the brain protein phosphorylated tau-217 (p-tau217) in a subset of participants. They were also linked to a more adverse social exposome — a composite measure of lifetime socioeconomic factors such as education, food insecurity, financial stress and healthcare access — among both healthy controls and Alzheimer’s patients, indicating that cumulative social hardship may be reflected in how old a person sounds.
Ibáñez highlighted the broader significance of these social findings. “I was particularly struck that speech reflected not only brain and molecular aging but also the social exposome,” he said. “It suggests that the cumulative environments people live in may leave measurable traces in something as everyday as speech.”
The study found moderate-to-strong correlations between the speech age gaps and established brain and epigenetic clocks. Brain-age gaps derived from structural MRI gray matter volume and resting-state functional MRI showed correlations of 0.42 and 0.49, respectively, with the speech clock, and a combined brain clock yielded a 0.53 correlation. Epigenetic clocks showed weaker associations, around 0.2, but still present, particularly among healthy adults and those with Alzheimer’s disease.
When examining brain networks, the researchers observed that the speech clock aligned more closely with whole-brain aging than with most individual networks, with only the language network and the salience network showing modest links. “One striking finding was how strongly the speech clock aligned with brain-age measures across several dementia syndromes, despite being derived solely from speech,” García noted. “We were also intrigued that whole-brain aging measures related more strongly to speech aging than most individual functional networks did. That suggests speech may reflect the coordinated state of distributed brain systems rather than the condition of a single language circuit.” Ibáñez added, “Speech appears to contain much more information about aging than we usually assume. Our speech-derived aging measure was associated not only with cognitive and clinical impairment, but also with brain aging, epigenetic aging, p-tau217, and social exposures.”
However, the authors acknowledge limitations. The study involved Spanish-speaking populations across five countries and relied on structured tasks, so results may not translate to other languages or casual conversations. The cross-sectional design cannot determine whether an older-sounding voice predicts future cognitive decline or tracks the speed of aging over time. There were also relatively small numbers in some subgroups, and not all participants had blood, brain-imaging or epigenetic data.
Commenting independently, Mai-Carmen Requena-Komuro described the speech-age gaps as a novel marker of ageing that can distinguish healthy elderly people from dementia patients in a large multi-country Spanish-speaking cohort. She cautioned, however, that certain speech changes are characteristic of language-variant forms of frontotemporal dementia and urged restraint in using SAG as a universal measure of degraded speech or disease progression. “Correlation does not mean causation,” she said, emphasising the need for longitudinal studies.
The researchers emphasise that the speech clock should complement, not replace, established clinical assessments. They hope to extend their work across languages and naturalistic speech and to explore whether lifestyle and clinical interventions can shift speech aging in a healthier direction. They also aim to validate the models in diverse populations where access to MRI, PET and molecular biomarkers remains limited. “Speech-based approaches could be particularly valuable in settings where MRI, PET, or sophisticated molecular biomarkers are difficult to access at scale,” Ibáñez said.
The study, Speech clocks decode dementia phenotypes, social exposome, and biological aging, lists a broad team of researchers from institutions across Latin America and beyond.
