The Way You Say "Um" Could Be Your Brain's Early Warning System for Dementia.

Think about the last conversation you had.

Not what you said. Not the words themselves or the point you were making. Think about the texture of it. The moment a word you knew perfectly well took an extra half-second to surface. The "um" that appeared not because you lacked the answer, but because somewhere between knowing it and saying it, there was a gap — small, unremarkable, the kind of thing you would never think twice about.

You didn't think twice about it. Neither did the person you were talking to.

But scientists are now paying very close attention to exactly those moments. Not because a single "um" means anything on its own. But because the pattern of those micro-events in speech — their frequency, their duration, where they appear in a sentence, and how they shift subtly over months and years — is, mounting evidence suggests, something remarkable: a real-time biological broadcast of how well your brain is functioning. And in some cases, a warning signal that arrives up to a decade before any clinical diagnosis of dementia.

The window for intervention in Alzheimer's disease is not the year of diagnosis. It is the years before it. And for the first time in the history of neuroscience, the tools to see inside that window are here — and they are listening to you talk.

What the Brain Scans Found

In 2025, researchers at Baycrest, the University of Toronto, and York University conducted a study that was deceptively simple in design and significant in implication.

Participants were shown detailed images and asked to describe them in their own words. Their speech was recorded. They also completed standard cognitive tests measuring executive function — the group of mental abilities that includes memory, planning, attention, processing speed, and flexible thinking. These are the capacities that neurodegenerative disease compromises first and most severely.

The researchers then fed the recordings into an AI system capable of analyzing hundreds of fine-grained timing and fluency markers simultaneously: how often people paused, how long those pauses lasted, how frequently they used filler words like "um" and "uh," and precisely how long it took them to retrieve specific words. The results, published in the Journal of Speech, Language, and Hearing Research, provided some of the strongest evidence yet that natural speech patterns are closely connected to executive function — even after controlling for age, sex, and education.

How someone's speech flowed predicted how well they performed on cognitive tests with a consistency that surprised the research team. Not what they said. How they said it.

The message, in the words of the lead researchers, was clear: "Speech timing and fluency may offer a practical window into cognitive health, even in individuals who haven't shown obvious signs of decline."

Why "Um" Is Not Just a Habit

To understand why filler words carry neurological information, you first need to understand what speech actually demands of the brain — because fluent conversation is, examined closely, one of the most computationally intensive acts a human being performs.

When you speak, your brain simultaneously manages phonological encoding, syntactic construction, lexical retrieval, and semantic monitoring — all in parallel, all in fractions of a second, coordinated across the prefrontal cortex, temporal lobes, basal ganglia, and cerebellum. It is an extraordinary real-time feat, performed so automatically that its complexity is invisible.

A filler word is not a failure in this system. It is the system reporting a processing delay. Research has established that "um" specifically tends to precede longer delays and is associated with deeper lexical retrieval difficulty — the brain reaching for a word that is taking longer than usual to arrive. "Uh" tends to precede shorter delays. These are not interchangeable verbal tics. They are functionally distinct signals, each carrying specific information about where in the speech production pipeline the bottleneck occurred.

Silent pauses carry different information again. A pause in the middle of a sentence — particularly in a grammatically unexpected location — signals that something in the planning or execution process stalled in a way the speaker did not anticipate. The brain reached for something and had to wait.

The executive functions linked to these patterns — working memory, processing speed, cognitive flexibility — are precisely the functions that degrade earliest in Alzheimer's disease. The frontal and temporal circuits that coordinate rapid, fluent speech are among the first to accumulate amyloid plaques and tau tangles. The degradation of those circuits does not first manifest as a missed name or a lost key. It manifests, in the brain's own language, as infinitesimal delays in the most cognitively demanding real-time task a human being performs.

The AI is not inventing meaning in these signals. It is decoding a message the brain has been broadcasting for years.

The Writer Who Left a Decade of Clues

Perhaps no case study in this field is more quietly devastating than what researchers at Cardiff and Loughborough Universities found in the novels of Sir Terry Pratchett.

Pratchett — beloved author of the Discworld series — was diagnosed with Posterior Cortical Atrophy, a form of dementia caused by Alzheimer's disease, in 2007. He died in 2015. Researchers decided to ask: could linguistic analysis of his 33 Discworld novels detect signs of his dementia before his clinical diagnosis?

They measured lexical diversity — the richness and variety of vocabulary — across each novel, tracking specifically how the range of nouns and adjectives evolved throughout his career. The results were striking. A significant decrease in lexical diversity was observed in Pratchett's later works, with total word count increasing while vocabulary variety decreased — a shift toward simpler, more repetitive language structures.

This shift occurred approximately ten years before his formal diagnosis.

The changes were invisible to readers. They were not detectable in any narrative, stylistic, or quality assessment of the books. They were only visible when the language was measured mathematically — and when measured, they were unambiguous.

"Language deficits may be observed many years before a formal diagnosis," the researchers concluded, "indicating that Alzheimer's disease has a long preclinical period."

In Pratchett's case, that period was nearly a decade. The brain was already changing. The language changed with it. Nobody was counting the adjectives.

What AI Hears That Humans Cannot

The Baycrest and Pratchett studies sit within a research landscape that has been expanding rapidly — and the technology driving that expansion is transforming what is detectable.

Researchers at Louisiana State University found that longer pauses in speech, especially during memory tasks, can reveal cognitive decline before other symptoms appear. Crucially, the signal was present even when participants ultimately gave the correct answer — the pause before accuracy, not the error itself, was the informative event. Standard cognitive testing scores what you get right or wrong. It does not measure the time it took to retrieve the answer. AI-powered speech analysis does — with millisecond precision, across hundreds of acoustic and linguistic features simultaneously.

Research published in npj Dementia using AI analysis of voice recordings from cognitive impairment studies found that a deep learning model achieved an AUC of 0.988 for detecting mild cognitive impairment from speech alone. AUC values approaching 1.0 represent near-perfect diagnostic accuracy. These are numbers that rival the diagnostic accuracy of PET scans and cerebrospinal fluid biomarker tests — derived not from brain imaging or invasive procedures, but from voice recordings.

A Boston University-led research team found that AI speech analysis predicted the progression from mild cognitive impairment to Alzheimer's disease within six years with over 78% accuracy. Prediction, not just detection. The distinction matters enormously — because the window in which disease-modifying interventions have the greatest potential is not the moment of diagnosis. It is the years before it, when the biological process is advancing but the clinical threshold has not yet been crossed.

Research tracking participants who carried genetic risk factors for frontotemporal degeneration found notable speech differences between at-risk individuals and non-carriers years before any clinical symptoms emerged — detectable across thirty measured speech features including rate, pause patterns, filler word frequency, and word repetition.

The convergence across these independent studies is striking: AI-powered speech analysis can detect the early cognitive changes associated with dementia earlier, less invasively, and in many cases more accurately than the tools currently in widespread clinical use.

Why Speech Fails Before Memory Does

There is a question this research consistently raises: why does speech degrade before memory? Why is the linguistic signal of Alzheimer's visible years before the clinical memory symptoms that finally bring people to a physician?

The answer is neuroanatomical. In its earliest stages, dementia can affect attention, perception, and language before memory problems become obvious. These early changes are difficult to detect because they are gradual and easily mistaken for stress, aging, or normal variation in behavior.

The prefrontal and temporal circuits that coordinate complex language — rapid lexical retrieval, syntactic planning, semantic monitoring — are affected early in the Alzheimer's cascade, before the hippocampal damage that produces the memory failures that define the disease in public awareness. Speech is not merely correlated with cognition. It is one of the most demanding cognitive acts the brain performs, and its demands are concentrated precisely in the brain regions that Alzheimer's pathology compromises first.

The "um" that resolves a half-second later than it used to. The sentence that pauses in a grammatically unexpected place. The word that arrives fractionally slower than before. These are not communication failures. They are the brain reporting that the processing infrastructure behind them is under pressure it was not under a year ago.

This is why the speech signal precedes the memory signal. Language becomes the brain's first legible distress call — legible, that is, to an instrument sophisticated enough to read it.

The Bottleneck Speech Could Break

Worldwide, 55 million people are currently living with dementia. That number is projected to reach 152 million by 2050. An estimated 75% of people with dementia globally are never formally diagnosed. Every 65 seconds, an American develops Alzheimer's disease.

The barriers to earlier diagnosis are well understood: specialist scarcity, geographic inequity, the cost and invasiveness of gold-standard biomarker tests, the time burden of comprehensive neuropsychological assessment. A speech analysis tool requires none of these. It requires a microphone and a few minutes of natural conversation — describing an image, recalling a story, answering open questions. The data collection is performable in a primary care office, a telehealth call, or eventually a smartphone application tracking longitudinal changes in a person's speech over months and years.

Practical and scalable — the specific combination medicine needs for a population-level dementia screening tool. Brain scans do not scale. Lumbar punctures do not scale. Conversations, recorded and analyzed, scale to every language, every geography, and every economic stratum that has access to a phone.

What to Do With This

The most important practical implication of this research is not about "um."

It is about baseline.

The reason longitudinal speech analysis can detect decline is that it measures change — and change requires a reference point. The earlier you establish your personal baseline, the more sensitive any future monitoring becomes. Researchers argue that speech-based cognitive monitoring should begin not at 65 or 70 but in middle age, when the preclinical phase of Alzheimer's has, for many people who will eventually develop the disease, already begun. Pathological changes in Alzheimer's disease begin up to decades before diagnosis.

If you have a family history of Alzheimer's or related dementias, raise the topic of early cognitive monitoring with your physician. If you notice — or people close to you notice — persistent changes in your speech fluency, word-finding, or conversational rhythm that represent a departure from your norm, take that signal seriously enough to investigate.

The research is not saying that using "um" means you are developing dementia. It is saying that a meaningful, sustained change in your own baseline — more pauses, slower retrieval, more frequent filler words than before — may carry information worth bringing to a doctor.

Your brain has been speaking in this language for your entire life.

The Story the Brain Has Always Been Telling

There is something deeply human in the central finding of this research. The brain, in the earliest stages of a disease that will eventually steal language entirely, does not go silent. It speaks. It signals. It broadcasts, in the texture and timing of ordinary conversation, information about its own condition — a message, in many cases, a decade before anyone thought to listen.

Terry Pratchett wrote 33 novels. In the adjectives he stopped reaching for, in the vocabulary that narrowed while his word count grew, he was writing — unknowingly, without intention — the story of what was beginning to happen to his brain. The story was there. The instruments to read it had not yet arrived.

They are arriving now.

The question is whether medicine will build the infrastructure to listen at scale — to make the technology available to the millions whose brains are, right now, sending a message in the only language it has always had.

The brain writes its autobiography in the way it speaks.

We are finally learning to read it.