How an AI turns a spoken story into an ageing clock

Name animals for a minute. Retell a story, then recall it after a delay. Tasks like these supplied recordings for a speech clock described in Science Advances on 30 September 2026.[1]

The study included 2,928 Spanish speakers across five Latin American countries. Seven families of speech measurements fed an age-prediction model. Its reported cross-validation mean absolute error was 9.04 years: an error in estimating calendar age, not a measure of diagnostic accuracy.[1]

The input reaches beyond how old a voice sounds to a listener. Trinity College Dublin’s account describes measurements of pauses, pitch, vocabulary, semantic precision and the organisation of speech. The researchers’ interest spans both the sound produced and the language assembled. That makes a remembered story a particularly rich object: it has words, timing and a sequence of events.[3]

What the remainder means

The underlying idea already has a history in brain imaging. A 2019 study by Stephen Smith and colleagues examined the difference between age inferred from scans and a person’s calendar age. The purpose is to investigate departures from an age-related pattern. A prediction can be imperfect as a birthday estimate while its departures still carry information about people.[2]

For an illustrative calculation, imagine a 70-year-old whose recording receives an estimated age of 76. Subtraction produces a raw gap of six years. That does not establish six extra years of biological wear, or tell us how quickly the person reached that point. It describes a position relative to the model’s learned pattern.[2]

There is a statistical catch: age-prediction models can pull estimates towards the middle of the age range, making younger people appear older and older people younger. Smith’s study explains why that bias can manufacture or obscure associations. Correcting the gap matters as much as predicting age accurately.[2]

The speech researchers corrected age-related bias using training data. Patient groups had larger gaps than healthy controls. Differences ranged from 0.33 to 0.76 on Cliff’s delta, a measure running from −1 to +1; positive values here indicate higher patient gaps. The largest difference involved language-dominant frontotemporal dementia.[1]

Among Alzheimer’s participants with blood data, gaps correlated modestly (r=0.18) with p-tau217, a blood marker of Alzheimer’s pathology. These cross-sectional comparisons establish neither individual ageing speed nor future dementia risk.[1]

A recording that could be repeated

Senior author Agustin Ibanez describes the prospect of combining speech with other biomarkers, potentially complementing costly measurements. The attraction is practical: recording someone repeatedly could be far more accessible than repeatedly arranging advanced imaging. The team says further work must include other languages, everyday speaking conditions and follow-up over time before clinical use.[3]

A single story can show where someone sits in the clock’s pattern. The more revealing recording may eventually be a second telling, under comparable conditions, months or years later.[3]

Report an error or suggest a correction