What speech can tell us.
And what it can’t.
Speech and language are promising research signals for cognitive change. They are not a stand-alone diagnosis of Alzheimer’s disease or another dementia. This brief explains three approaches and what this pilot actually measures.
Early detection and younger onset
“Early onset” can mean two different things. In clinical usage, younger-onset Alzheimer’s means symptoms begin before age 65; “early detection” means finding changes in an early disease stage at any age. A person with younger-onset disease may be in an early, middle or late stage. The speech studies below mostly examine cognitive impairment or Alzheimer’s in older cohorts, so their findings should not be assumed to generalize to younger-onset disease. Terminology ↗ · Early- versus late-onset language study ↗
Why idea density mattered
The 1996 Nun Study linked idea density in autobiographical writing from early adulthood with cognition and Alzheimer’s pathology decades later. It was a striking population finding, but it does not mean counting ideas in a single sentence can tell an individual whether they have dementia. Proper propositional idea density involves identifying distinct propositions, a task our pilot does not perform. Read the original study ↗
Three approaches worth exploring
1. Explainable language features
Measure semantic content, vocabulary, word finding, syntax, repetitions and discourse from a standard prompt. A 2025 study of 336 DementiaBank participants found differences in lexical, structural, syntactic and discourse features across cognitive groups. This is a practical starting point for a transparent personal journal, but simple features are sensitive to task and education. Study ↗
2. Acoustic and timing features
Measure pauses, speech rate, fluency and other voice features. A 2026 multisite study reported that speech features in a brief dual task improved discrimination of mild cognitive impairment compared with demographics alone (AUC 0.79 versus 0.56 in 319 usable recordings). That result applies to the study protocol and should not be transferred to arbitrary uploads. Study ↗
3. Combined and longitudinal models
Pair audio and transcripts, then compare repeated, standardized tasks. A 2026 longitudinal dual-modal study reported improved performance over single-visit models in two Chinese-language datasets. A 2025 study connected speech-derived features from cognitive tasks with later cognitive and brain changes. This direction is promising but needs independent, diverse validation. Longitudinal study ↗ · Cognitive trajectories ↗
Relevant projects and datasets
- DementiaBank is a major speech and transcript research resource. Most recordings require controlled researcher access.
- ADReSS and ADReSSo are benchmark challenges for dementia recognition and cognitive score or decline prediction from speech.
- ADscreen is a research speech processing screening system.
- Winterlight Labs research has studied automated detection of progressive speech changes in early Alzheimer’s disease.
- Framingham speech research has studied language model based prediction of future Alzheimer’s progression.
What Dementia measures today
For English text, Dementia shows word count, sentence length, content word share (a simple stopword-based proxy), vocabulary variety within 50-word blocks, filler share and repeated three-word phrases. For audio, it detects active speech and pauses from signal energy. These are descriptive measurements. They are not validated Alzheimer’s biomarkers, proposition counts, cognitive scores, or clinical risk estimates. Audio is not automatically transcribed; a supplied transcript adds text analysis.
The app saves report measures and dates, but discards uploaded media and submitted text after analysis. Comparisons use the same task and input type. Different microphones, speech topics, language backgrounds, hearing, sleep, mood, illness and medication can change the results.
A 2026 systematic review of 63 AI speech studies found promising results for distinguishing diagnosed Alzheimer’s disease from healthy controls, but weaker detection of mild cognitive impairment and limited external validation. Clinical diagnosis requires medical history, functional and cognitive assessment, and sometimes imaging or biomarker tests. Systematic review ↗ · Clinical evaluation overview ↗
Why the age reference is a separate task
There is no defensible age percentile for an arbitrary story, document, or audio upload: the prompt, sample length, language, schooling, and recording conditions differ from published tests. Dementia therefore offers an optional one-minute spoken animal-naming check. The person can count manually or review names suggested from a voice transcript, then enters age and years of education. The app compares that count with broad percentile bands in Table 4 of a Canadian community sample. For some age and education combinations, the table has no usable estimate. These historical norms were collected under a specific protocol and do not establish a dementia probability. Tombaugh et al. (1999) ↗ · Animal naming administration and scoring ↗
Age and education both affect verbal fluency, and norms differ across languages and populations. A self-tallied result at home is only a rough reference, not a clinical test. Cross-linguistic review ↗
How to read the report colors
For uploaded audio and writing, the green, yellow and red labels are technical measurement guides. They indicate whether this prototype received enough audio activity or words to calculate its basic features, and whether the audio has many near-peak samples that could indicate clipping. The guides were chosen for the app; they are not published healthy-versus-dementia thresholds. The remaining cards show their mathematical scale or state that no validated range exists for an arbitrary task. After at least three comparable uploads, the report can also show the person’s own observed range from up to five earlier samples. It favors different-day visits when available; a same-day range describes variation between recordings rather than change over time. An outside value has no favorable or unfavorable direction by itself.
The one-minute animal-naming task has a different, published reference: red means below the 10th percentile, yellow means the 10th–24th percentile band, and green means the 25th percentile or higher in the selected age and education group from Tombaugh et al. (1999). These colors describe a reviewed or self-counted result’s position in that historical sample, not disease status.
Published pause findings cannot be applied directly to the upload cards. For example, one study used a picture-description task and counted pauses longer than 150 milliseconds, while this app estimates pauses longer than about 360 milliseconds from sound energy. Recording devices, noise and task also affect values. Current reviews call for task-specific, demographic and language-sensitive norms and external validation before clinical interpretation. Pause study ↗ · Speech testing review ↗
Population benchmarks and repeat timing
Yes, healthy-population speech reference data exist. A study of 290 healthy participants aged 15–90 doing picture-description tasks found that speech percentage tended to decrease and pause rate tended to increase with age. Another healthy-speaker study reported task-specific acoustic values. These are valuable research baselines, but neither provides a compatible percentile for this app’s free-form prompts and sound-energy detector. To offer a defensible population band for the uploaded audio metrics, the app would need a standardized task and a healthy reference cohort processed with the same algorithm, with enough people across age, education, language and recording-device groups. Healthy picture-description cohort ↗ · Healthy age-group comparison ↗ · Validation framework ↗
There is no validated optimal repeat interval for this pilot. In research, the SIDE-AD protocol collects samples every three months and DASH every two months. A repeated-speech methods study also sampled within one day and one week to understand ordinary variation, then repeated a session 8–11 weeks later. As a practical journal guide, collect two or three comparable recordings a few days apart to see short-term variation, then repeat about every two to three months under similar task, language, device, room and time-of-day conditions. This is an inference from research schedules, not a clinical monitoring recommendation. SIDE-AD ↗ · DASH ↗ · Repeat-speech protocol ↗
How the new spoken tasks work
In the animal-naming exercise, the browser records one minute of speech or accepts an uploaded recording. The server transcribes it with the official OpenAI Whisper base.en model running locally, or with OpenAI’s gpt-transcribe API if a server key is configured and the user consents to sending audio to OpenAI. A word-list parser proposes animal names, but the person reviews and corrects them before the unique count is saved. Automatic transcription may omit, substitute or invent names, so the unreviewed list is not used as a final score. OpenAI transcription guidance ↗ · Animal naming scoring ↗
The fixed picture-description exercise uses an original illustrated park scene. Its report uses a local pretrained Qwen2.5-1.5B-Instruct language model to compare the reviewed transcript with ten predefined visual facts. The model proposes which facts were explicitly described; the person reviews and may correct those marks before saving. It does not rely on keyword matching. This semantic scoring is experimental and can make mistakes. Healthy picture-description cohorts show why this is a useful standardized task type, but they did not use this scene or detector. The detail count therefore has no population percentile or clinical cutoff. Establishing one would require a healthy reference cohort describing this exact scene, processed by the same pipeline and stratified appropriately. Healthy picture-description cohort ↗
Before any clinical use
- Define a repeatable task and validate transcription and acoustic quality.
- Collect consented longitudinal data with clinical and, where appropriate, biomarker reference outcomes.
- Evaluate age, education, accent, language, sex, hearing and device effects; test on independent sites.
- Calibrate any model, measure false positives and negatives, and study whether the report changes care beneficially.
- Review medical-device, privacy and data governance requirements with qualified specialists.