Your voice could add something aging blood tests and scans miss

By Dr. Sanchari Sinha Dutta, Ph.D.Reviewed by Lauren HardakerSep 15 2026

Researchers tested whether a brief counting task could capture age-related physiological information that established aging clocks may miss.

Study: Sound of aging: large-scale evidence for a voice-based biological clock. Image Credit: elebeZoom / Shutterstock

Voice may serve as an accessible marker of functional aging, according to a new study published in the journal npj Aging.

Background

Aging is a heterogeneous process in which molecular, physiological, and functional damage accumulates differently across individuals over time. Chronological age refers to the amount of time a person has lived since birth and does not fully capture individual aging trajectories.

Biological-age measures estimate aspects of a person's molecular, cellular, and physiological state using epigenetic, proteomic, and imaging data and have improved age-related risk prediction. Many of these methods are expensive or invasive, making large-scale implementation challenging. These factors highlight the need for widely deployable, non-invasive biomarkers that can link cellular damage to functional decline.

Researchers at the Weizmann Institute of Science in Israel investigated whether the human voice could serve as a candidate biomarker of functional aging.

Human speech production is a complex motor process requiring coordinated actions of multiple physiological systems. With aging, these systems undergo well-documented structural and functional declines that directly affect voice quality.

Presbylaryngis is characterized by age-related vocal fold atrophy, bowing, and reduced mucosal wave amplitude, which impair vibratory efficiency and alter voice quality. Systemic age-related changes, including sarcopenia, reduced connective-tissue elasticity, and neuromotor decline, can further affect voice production.

These age-related changes are reflected in measurable features of human speech, including elevated spectral noise, reduced cepstral peak prominence and harmonics-to-noise ratio, altered spectral slopes, and increased fundamental-frequency variability.

Human listeners can naturally use these vocal changes to estimate a speaker's age with a mean absolute error of 5 to 10 years, providing a reference point for evaluating the performance of computational models.

In the current study, researchers hypothesized that deep-learning speech representations would predict age more accurately than the four prespecified engineered acoustic feature sets and would encode established acoustic dimensions of vocal aging.

They analyzed standardized 30-second voice recordings collected from 6,979 Israeli adults aged 40–70 years, in which participants counted aloud from 1 to 30 at a comfortable pace. They trained sex-stratified models using deep-learning speech embeddings to predict chronological age exclusively from vocal features, and termed the resulting prediction "Voice Age".

They also assessed Voice Age acceleration as the residual from a sex-specific linear regression of Voice Age on chronological age, indicating whether a speaker's voice sounded older or younger than expected based on the fitted age trend.

Key findings

The sex-stratified models explained 54% of the variance in chronological age among female participants and 44% among male participants. For females and males, the mean absolute errors for age predictions were 3.95 years and 4.41 years, respectively. These errors are lower than those reported in human-listener studies.

In the single-modality comparison, Voice Age showed the second-highest predictive performance among nine single-modality age models and correlated only partially with other established biological aging clocks. This ranking was descriptive because the models were evaluated in modality-specific cohorts of different sizes.

Adding Voice Age to eight molecular, imaging, physiological, and lifestyle models improved age prediction over the stronger individual input in 159 of 160 evaluations across sexes and modalities. When combined with mass-spectrometry metabolomics, R² reached 65% in females and 52% in males.

In both males and females, Voice Age acceleration was associated with greater adiposity, sleep-disordered breathing, lower nocturnal oxygen saturation, and differences in hepatic imaging measures.

Voice Age acceleration also showed sex-specific associations with grip strength, cardiometabolic measures, and bone health measures. Based on these findings, researchers suggest that Voice Age acceleration should be interpreted as a sex-dependent readout integrating body composition, sleep, hepatic, skeletal, and functional physiology.

In males, lower cholesterol and triglycerides in the older-sounding group ran counter to a simple adverse cardiometabolic gradient, so Voice Age acceleration should not be treated as a uniformly directional cardiometabolic risk score.

Study significance

The study shows that a 30-second voice recording contains a reproducible age-related signal and captures physiological information partly distinct from established aging clocks. Voice Age also added predictive information to individual molecular and physiological models.

Deep-learning speech embeddings used to train the models linearly encoded established acoustic dimensions of voice aging, indicating that the embeddings preserve physiologically meaningful vocal structure. The analysis did not identify which features drove age prediction or establish performance across new recording environments.

These findings support Voice Age as a candidate functional aging biomarker that could be inexpensive to collect. Since a voice sample can be collected easily and repeatedly, a prospectively validated Voice Age measure could provide an inexpensive first layer of risk stratification and longitudinal monitoring and help identify individuals who may benefit from more intensive physiological or molecular assessment.

Because of the cross-sectional design, Voice Age acceleration reflects deviation from the sex-specific age trend at one time point and does not establish an individual's aging trajectory. The observed associations do not establish causality, future disease risk, or screening performance.

The exploratory phenome-wide analyses were not adjusted for potential confounders beyond linear age residualization, and test-retest reliability across sessions, devices, and acoustic environments remains to be established.

The study population included only Hebrew-speaking Israeli adults aged 40–70 years from a single site, while WavLM-Large was pre-trained primarily on English speech. No cross-language transfer analysis was performed in the present study. These factors leave generalizability beyond this age range and to other languages, sites, recording conditions, and clinical populations uncertain.