A 2025 meta-analysis of 54 studies found that speech-based tools separate mild cognitive impairment (MCI) from healthy ageing with a pooled area under the curve of 78.0%, which means a randomly chosen person with MCI scores as more likely impaired than a randomly chosen healthy person about 78 times out of 100. The same review reports 80.0% sensitivity and 77.0% specificity (Jafari et al., Age and Ageing 2025). Shared research challenges show how far the field has moved on a fixed dataset: the 2020 ADReSS winner reached 89.58% accuracy against a 75% baseline, while a Framingham Heart Study model built on 1,084 participants reached 92.6% AUC for normal cognition versus dementia but 74.4% for the harder normal-versus-MCI task. Meanwhile, only 10.6% of US Medicare beneficiaries who met survey criteria for MCI had a recorded diagnosis, which is the gap voice biomarkers are meant to close. We aggregated data from peer-reviewed meta-analyses in Age and Ageing and the Journal of Prevention of Alzheimer’s Disease, the ADReSS challenge papers, the Framingham Heart Study, TalkBank DementiaBank, the Alzheimer’s Association, WHO, the FDA and the Global CEO Initiative on Alzheimer’s Disease. For the hearing side of the same ageing-brain story, see our separate roundup on hearing loss and dementia risk.
This page reports research findings only. It is not medical advice, and none of the tools described here replace a clinical evaluation.
TL;DR
- Speech biomarkers detect MCI with a pooled 78.0% AUC and 80.0% accuracy across 21 studies (Jafari et al., Age and Ageing 2025).
- People with Alzheimer’s disease pause about 1.20 standard deviations longer than healthy controls (Cohen et al., Journal of the International Neuropsychological Society 2026).
- The ADReSS 2020 challenge winner reached 89.58% accuracy on a 48-person test set (Luz et al., IEEE Open Journal of Signal Processing 2024).
- Models trained on English and tested on Greek reached a top accuracy of 87.0% in ADReSS-M (Luz et al., IEEE Open Journal of Signal Processing 2024).
- A Framingham language model hit 92.6% AUC for normal vs dementia, but 74.4% for normal vs MCI (Amini et al., Alzheimer’s and Dementia 2023).
- Speech from 166 Framingham participants with MCI predicted progression to Alzheimer’s within 6 years with 78.5% AUC (Amini et al., Alzheimer’s and Dementia 2024).
- DementiaBank lists 23 corpora in 7 languages, 13 of them English (TalkBank DementiaBank 2026).
- 84% of 51 speech-AI studies reviewed had no external validation (de la Fuente Garcia et al., Journal of Alzheimer’s Disease 2020).
- Average MCI detection in US primary care was 0.08 of expected cases across 226,756 clinicians (Liu et al., Journal of Prevention of Alzheimer’s Disease 2024).
- Expert standards ask detection tools for at least 80% sensitivity and 85% specificity (Global CEO Initiative on Alzheimer’s Disease, Alzheimer’s and Dementia 2025).
- 7.4 million Americans aged 65 and older live with Alzheimer’s dementia in 2026 (Alzheimer’s Association, 2026 Alzheimer’s Disease Facts and Figures).
1. Pooled Detection Accuracy: What the Meta-Analyses Say
The headline number is respectable but not decisive. A pooled specificity of 77.0% means roughly 23 of every 100 cognitively healthy people tested would be flagged in error (100% minus 77%), and that matters far more in a primary care waiting room than in a research lab where half the sample is impaired by design. The pooled sensitivity of 80.0% sits exactly at the floor that expert standards set for a detection tool, while the specificity sits below the 85% floor (see section 7).
The Age and Ageing meta-analysis screened 4,432 records and kept 54 articles, 35 of which fed the pooled estimates. Its authors list the limitations plainly: speech tasks, features and machine learning models varied widely, 25 of the 54 studies had fewer than 30 participants per group, and models trained mainly on English speakers may not carry over to other languages, dialects or education levels.
| Metric | Value | Source |
|---|---|---|
| Pooled accuracy, speech-based tools, MCI vs cognitively unimpaired adults | 80.0% (95% CI 70.0% to 89.0%, 21 studies) | Jafari et al., Age and Ageing 2025 |
| Pooled area under the curve (AUC), MCI vs cognitively unimpaired adults | 78.0% (95% CI 70.0% to 86.0%) | Jafari et al., Age and Ageing 2025 |
| Pooled sensitivity and pooled specificity | 80.0% (22 studies) and 77.0% (15 studies) | Jafari et al., Age and Ageing 2025 |
| Studies pooled for pause length vs for speaking rate | 13 vs 8 studies | Cohen et al., Journal of the International Neuropsychological Society 2026 |
| Pause length, people with Alzheimer’s vs healthy controls | About 1.20 SD longer (MCI: about 0.62 SD longer) | Cohen et al., Journal of the International Neuropsychological Society 2026 |
| Speech rate difference, people with Alzheimer’s vs healthy controls | 0.66 SD (MCI: 0.27 SD) | Cohen et al., Journal of the International Neuropsychological Society 2026 |
| Acoustic parameters that differed significantly, Alzheimer’s vs healthy controls | 5 of 7 parameters, 11 studies | Saeedi et al., Journal of Prevention of Alzheimer’s Disease 2024 |
| Speech samples pooled in the acoustic meta-analysis | 1,000 (386 Alzheimer’s, 614 healthy controls) | Saeedi et al., Journal of Prevention of Alzheimer’s Disease 2024 |
Context note: SD stands for standard deviation, a way to express a group difference independent of the unit measured. A difference of about 1.2 SD is large; 0.27 SD is small. The pause meta-analysis reported substantial heterogeneity across its studies, meaning results varied more than chance alone would explain. The acoustic meta-analysis adds a caution of its own: slowed speech also appears in other neurodegenerative conditions such as Parkinson’s disease and ALS, so a slow speaker is not automatically an Alzheimer’s case. Our roundup on Parkinson’s speech and voice covers that overlap from the treatment side.
2. The ADReSS Benchmarks: Shared Challenges on a Fixed Dataset
Shared challenges are the closest thing this field has to a common yardstick. Every team gets the same recordings of the Cookie Theft picture description (a standard line drawing of a kitchen scene that people describe aloud) and is scored on the same hidden test set. The jump from a 75% baseline to an 89.58% winning accuracy in ADReSS 2020 shows what modern language models can extract from a single picture description.
The fine print is the test set size. The ADReSS 2020 paper held out 48 speakers, so each misclassified person moves accuracy by about 2.1 percentage points (1 of 48 = 2.08%). The ADReSSo 2021 challenge removed manual transcripts and asked teams to work from audio alone, and the ADReSS-M overview went further, training on English and testing on Greek.
| Metric | Value | Source |
|---|---|---|
| ADReSS 2020 dataset, people with Alzheimer’s vs people without | 156 speakers (78 and 78), split 108 training and 48 test | Luz et al., ADReSS Challenge, Interspeech 2020 |
| ADReSS 2020 baseline accuracy on the test set, transcript-based vs audio-based model | 75% vs 62.5% | Luz et al., ADReSS Challenge, Interspeech 2020 |
| ADReSS 2020 winning team accuracy | 89.58% | Luz et al., IEEE Open Journal of Signal Processing 2024 |
| ADReSSo 2021 baseline, audio-only Alzheimer’s detection | 78.87% accuracy | Luz et al., ADReSSo Challenge, Interspeech 2021 |
| ADReSSo 2021 baseline, predicting cognitive decline over two years | 68.75% accuracy | Luz et al., ADReSSo Challenge, Interspeech 2021 |
| ADReSS-M 2023 data, English training set vs Greek test set | 225 speakers (110 Alzheimer’s) vs 46 speakers (22 Alzheimer’s) | Luz et al., IEEE Open Journal of Signal Processing 2024 |
| ADReSS-M 2023 top team, Alzheimer’s detection accuracy and MMSE error | 87.0% accuracy, RMSE 3.727 | Luz et al., IEEE Open Journal of Signal Processing 2024 |
| ADReSS-M 2023 participation | 24 teams from 14 countries, 17 submitted to both tasks | Luz et al., IEEE Open Journal of Signal Processing 2024 |
Context note: MMSE is the Mini-Mental State Examination, a 30-point paper-and-pencil cognitive test, and RMSE (root mean square error) is the typical size of the gap between a model’s predicted score and the real one, in MMSE points. The 2021 decline task defined decline as an MMSE drop of 5 points or more between baseline and year two. The organizers themselves warn of over testing: when a large community tunes models on one small dataset, results may reflect parameter choices more than any general link between language and cognition.
3. Speech Datasets: The Data Bottleneck Behind the Models
Most of the accuracy figures above trace back to a small number of corpora. DementiaBank, the main shared repository, lists 23 corpora, and 13 of them are in English, while the remaining 10 are spread across German, Greek, Korean, Mandarin, Spanish and Taiwanese. That concentration is the practical reason reviewers keep warning that models may not travel across languages.
Access is deliberately gated. The DementiaBank repository is password protected and restricted to members of its consortium, students need a faculty sponsor, and the project is funded by the US National Institute on Aging. The site also carries a notice that the repository is under review for potential modification in compliance with Administration directives, a reminder that a single policy change can affect the field’s reference data. Speech recognition errors matter here too, because many pipelines start from automatic transcripts; our roundup on speech recognition accent bias and word error rates shows how unevenly that first step performs.
| Metric | Value | Source |
|---|---|---|
| DementiaBank corpora and languages listed | 23 corpora in 7 languages | TalkBank DementiaBank access index 2026 |
| English corpora as a share of all DementiaBank corpora | 13 of 23 | TalkBank DementiaBank access index 2026 |
| Pitt corpus participants: controls, dementia, unknown diagnosis | 104, 208 and 85 | TalkBank DementiaBank Pitt corpus page 2026 |
| Spanish Ivanova corpus: Alzheimer’s, healthy controls, MCI | 74, 197 and 90 | TalkBank DementiaBank access index 2026 |
| Shared challenges listed on the DementiaBank site | 6 (including ADReSS 2020, ADReSSo 2021, ADReSS-M, TAUKADIAL-2024) | TalkBank DementiaBank 2026 |
| Speech-AI studies with no external validation | 84% of 51 studies | de la Fuente Garcia et al., Journal of Alzheimer’s Disease 2020 |
| Speech-AI studies that used class-balanced data (equal numbers of impaired and healthy speakers) | 39% (20 of 51 studies) | de la Fuente Garcia et al., Journal of Alzheimer’s Disease 2020 |
| Speech-AI studies trained on fewer than 50 participants vs 100 or more | About 31% vs 27% | de la Fuente Garcia et al., Journal of Alzheimer’s Disease 2020 |
Most recent available data for the study-quality audit: de la Fuente Garcia et al., Journal of Alzheimer’s Disease 2020, covering research published from 2000 to 2019. The same review found that about 41% of studies were conducted in English, and that most studies did not report whether their data could be accessed by others. The Pitt corpus counts come from the Pitt corpus page, which cites the original 1994 University of Pittsburgh cohort description.
4. The Framingham Heart Study Voice Cohort and Other Long-Running Cohorts
Framingham is unusual because its recordings were made for routine neuropsychological testing, not for machine learning. Those interviews have been digitally recorded since 2005, so researchers can test speech models on real clinical conversations and then check them against what happened to the same people years later. On raw audio, a convolutional neural network separated normal cognition from dementia with an AUC of 0.805, and a later model working from automatic transcripts of 1,084 participants reached 92.6% on the same comparison.
The harder number is the early one. Normal versus MCI fell to 74.4% AUC in the language-based model, and the 6-year progression study reached 78.5% AUC on 166 people. The authors of that study note that moderate specificity would flag many stable MCI patients in error, that participants were predominantly White, and that the model has not been validated outside Framingham.
| Metric | Value | Source |
|---|---|---|
| Framingham raw-audio study: participants and recordings | 656 participants, 1,264 recordings (September 2005 to March 2020) | Xue et al., Alzheimer’s Research and Therapy 2021 |
| Average length of one recorded neuropsychological session | 73.13 minutes (range 8.43 to 210.82) | Xue et al., Alzheimer’s Research and Therapy 2021 |
| AUC, normal cognition vs dementia, convolutional network vs LSTM network | 0.805 vs 0.740 | Xue et al., Alzheimer’s Research and Therapy 2021 |
| Language model on transcripts, AUC for normal vs dementia (1,084 participants) | 92.6% | Amini et al., Alzheimer’s and Dementia 2023 |
| Same model, AUC for normal vs MCI | 74.4% | Amini et al., Alzheimer’s and Dementia 2023 |
| Predicting progression from MCI to Alzheimer’s within 6 years: participants and AUC | 166 (90 progressed, 76 stable), AUC 78.5% | Amini et al., Alzheimer’s and Dementia 2024 |
| Long Life Family Study recall recordings, normal cognition vs cognitive impairment | 598 vs 112 participants | Park et al., Journal of the International Neuropsychological Society 2026 |
| Linguistic score vs traditional scoring of delayed story recall (precision-recall AUC) | 0.77 vs 0.81 | Park et al., Journal of the International Neuropsychological Society 2026 |
Most recent available data for raw-audio Framingham results: Xue et al., Alzheimer’s Research and Therapy 2021. LSTM (long short-term memory) and convolutional networks are two common types of deep learning model. In the Long Life Family Study analysis, the gap between 0.77 and 0.81 was not statistically significant (p = 0.292), and the authors warn that a circularity issue, because diagnoses partly relied on the same recall test, likely inflated accuracy.
5. Pathology-Anchored Studies: Amyloid, Tau and the Preclinical Frontier
Most speech studies compare people with a clinical diagnosis against healthy volunteers. The more demanding test is whether speech tracks the underlying biology, such as amyloid (a protein that builds up as plaques in the Alzheimer’s brain, measured by PET scan or spinal fluid). In the largest study here, the speech-based AUC for amyloid positivity drops to 0.71 within cognitively unimpaired adults, against 0.82 among people who already have MCI or mild Alzheimer’s.
That pattern runs through the small pathology-anchored studies: signals are clearest once symptoms exist and fade toward the preclinical stage. A Spanish memory-clinic study found acoustic features outperformed standard neuropsychological tests at predicting amyloid status in MCI, but the sample was 52 people from a single clinic. A Dutch at-home study showed people will do the recordings (91.6% adherence) but its amyloid-related differences did not survive correction for multiple testing.
| Metric | Value | Source |
|---|---|---|
| AMYPRED study sample, amyloid positive vs negative | 200 participants (97 vs 103) | Fristed et al., Brain Communications 2022 |
| AUC for amyloid positivity: full sample vs cognitively unimpaired subgroup | 0.77 vs 0.71 | Fristed et al., Brain Communications 2022 |
| AUC for detecting MCI or mild Alzheimer’s from about 6.56 minutes of speech | 0.83 | Fristed et al., Brain Communications 2022 |
| Amyloid status in MCI (52 patients, 30 amyloid positive): acoustic model vs neuropsychological tests | AUC 0.79 vs 0.66 | Garcia-Gutierrez et al., Frontiers in Neuroscience 2023 |
| At-home Dutch speech study, cognitively unimpaired adults: sample, amyloid positive, adherence | 50 adults, 23 amyloid positive (46%), 91.6% adherence | van den Berg et al., Alzheimer’s Research and Therapy 2024 |
| Longer pauses before adjectives and adverbs, amyloid- and tau-positive vs matched negative unimpaired adults | 26 vs 26 participants, p = 0.016 | Hale et al., Journal of Alzheimer’s Disease 2025 |
| MCI vs healthy controls: speech metrics vs standard cognitive measures | AUC 0.80 vs 0.71 | Garcia, Alzheimer’s and Dementia 2025 (conference abstract) |
Context note: the AMYPRED study was funded by Novoic Ltd, a speech-analysis company that employed several of its authors, and its sample was 193 white, 3 Asian and 4 Black or African American participants. Most recent available data for that design: Fristed et al., 2022. The Spanish memory-clinic study and the Dutch home study are both cross-sectional, so neither shows that speech changes track amyloid over time. The 2025 conference abstract reports AUC 0.89 for both word-property metrics and neuroimaging features in Alzheimer’s vs controls, a result available only in abstract form.
6. The Detection Gap: Disease Burden and Missed Diagnoses
The case for a cheap, scalable screen rests on two numbers: how many people are affected and how many are missed. Among 226,756 US primary care clinicians, the average MCI detection rate was 0.08 of the expected caseload, and only 0.1% of clinicians and practices fell within the expected range (Liu et al., Journal of Prevention of Alzheimer’s Disease 2024). A 2026 analysis of national survey data linked to Medicare claims landed on a similar picture from a different direction: 10.6% of people who met MCI criteria had the diagnosis recorded.
The burden behind that gap is large on any measure. WHO attributes 60% to 70% of dementia cases to Alzheimer’s disease and ranks dementia the seventh leading cause of death worldwide, and the Alzheimer’s Association reports that 1 in 3 older Americans dies with Alzheimer’s or another dementia. A voice test cannot change those totals, but it could change how early people enter care.
| Metric | Value | Source |
|---|---|---|
| People living with dementia worldwide, and Alzheimer’s share of cases | 57 million (2021); Alzheimer’s 60% to 70% of cases | WHO Dementia Fact Sheet 2026 |
| Global cost of dementia | US$ 1.3 trillion (2019) | WHO Dementia Fact Sheet 2026 |
| Americans aged 65 and older living with Alzheimer’s dementia | 7.4 million (2026) | Alzheimer’s Association, 2026 Alzheimer’s Disease Facts and Figures |
| Unpaid caregivers and unpaid care hours, United States | Nearly 13 million caregivers, more than 19 billion hours (2025) | Alzheimer’s Association, 2026 Alzheimer’s Disease Facts and Figures |
| Average MCI detection rate in US primary care, 226,756 clinicians | 0.08 of expected cases | Liu et al., Journal of Prevention of Alzheimer’s Disease 2024 |
| Medicare beneficiaries meeting survey MCI criteria with a recorded MCI diagnosis | 10.6% (593 respondents, about 2.6 million beneficiaries weighted) | Couch et al., Alzheimer’s and Dementia DADM 2026 |
| Share of US people with Alzheimer’s and related dementias who may be undiagnosed | More than 60% | Global CEO Initiative on Alzheimer’s Disease, Alzheimer’s and Dementia 2025 |
| Adults who have discussed brain health with their doctor | 14% | Alzheimer’s Association, 2026 Alzheimer’s Disease Facts and Figures |
Sources: WHO dementia fact sheet (updated July 3, 2026), Alzheimer’s Association Facts and Figures, and the Medicare MCI diagnosis study. In the Couch study, women (OR 2.13, 95% CI 1.29 to 3.51) and people with a bachelor’s degree or higher (OR 2.21, 95% CI 1.22 to 4.02) were more likely to have a recorded diagnosis, while people who attended doctor visits alone were less likely (OR 0.54). The 0.08 detection rate compares recorded diagnoses with a model of expected cases, so it depends on that model’s assumptions.
7. Regulatory Status and the Performance Bar for Digital Tools
Regulators do not yet treat a voice recording as a diagnosis. In the United States, a computerized cognitive assessment aid is a Class II prescription device under 21 CFR 882.1470: it interprets a person’s current level of cognitive function but cannot determine whether a clinical diagnosis is present and is not intended for stand-alone use. None of the sources reviewed for this roundup documents an FDA clearance of a speech-only Alzheimer’s diagnostic.
Expert standards now put a number on good enough: at least 80% sensitivity and 85% specificity for a brief detection tool, according to the Global CEO Initiative on Alzheimer’s Disease. Set against those floors, the pooled speech result in section 1 meets the sensitivity bar and misses the specificity bar by 8 points (85% minus 77%). The same paper shows why specificity matters in screening: at 10% prevalence, even a tool that meets the floor produces a positive predictive value of only 37%. The CEOi recommendations do not mention speech or voice at all.
| Metric | Value | Source |
|---|---|---|
| US device class for computerized cognitive assessment aids | Class II (special controls), not for stand-alone diagnosis | 21 CFR 882.1470 (eCFR 2026) |
| Recommended minimum for a brief detection tool, sensitivity and specificity | 80% and 85%, about 3 to 5 minutes | Global CEO Initiative on Alzheimer’s Disease, Alzheimer’s and Dementia 2025 |
| Recommended minimum for a diagnostic aid, sensitivity for MCI and specificity | 85% and 90%, about 10 to 20 minutes | Global CEO Initiative on Alzheimer’s Disease, Alzheimer’s and Dementia 2025 |
| Positive and negative predictive value of a tool at the detection floor, 10% prevalence | 37% and 97% | Global CEO Initiative on Alzheimer’s Disease, Alzheimer’s and Dementia 2025 |
| First FDA-cleared Alzheimer’s blood test: agreement with amyloid PET or CSF, positive and negative results | 91.7% and 97.3% (499 samples) | FDA press announcement, May 16, 2025 |
| Share of blood test results that were indeterminate | Fewer than 20% of 499 patients | FDA press announcement, May 16, 2025 |
| FDA Breakthrough Device designations vs marketing authorizations, all device types | 1,320 vs 210 (as of June 30, 2026) | FDA Breakthrough Devices Program 2026 |
Context note: positive predictive value is the share of people with a positive result who truly have the condition. The FDA blood test announcement is a useful benchmark because it shows the evidence package a cleared Alzheimer’s test now carries: 499 samples checked against PET or spinal fluid, with the agency stating the test is not intended as a screening or stand-alone diagnostic test. The Breakthrough Devices Program figures cover every device category, not speech tools specifically.
Summary: Voice Biomarkers for Alzheimer’s and MCI by the Numbers
| Metric | Value | Source |
|---|---|---|
| Pooled AUC, speech-based tools, MCI vs cognitively unimpaired adults | 78.0% (95% CI 70.0% to 86.0%) | Jafari et al., Age and Ageing 2025 |
| Pause length, Alzheimer’s vs healthy controls | About 1.20 SD longer | Cohen et al., Journal of the International Neuropsychological Society 2026 |
| Acoustic parameters that differed significantly in Alzheimer’s | 5 of 7 (11 studies) | Saeedi et al., Journal of Prevention of Alzheimer’s Disease 2024 |
| ADReSS 2020 baseline vs winning accuracy | 75% vs 89.58% | Luz et al., Interspeech 2020; IEEE Open Journal of Signal Processing 2024 |
| ADReSSo 2021 audio-only baseline accuracy | 78.87% | Luz et al., Interspeech 2021 |
| ADReSS-M 2023 top accuracy, trained on English, tested on Greek | 87.0% | Luz et al., IEEE Open Journal of Signal Processing 2024 |
| DementiaBank corpora, total and in English | 23 and 13 | TalkBank DementiaBank 2026 |
| Speech-AI studies with no external validation | 84% of 51 | de la Fuente Garcia et al., Journal of Alzheimer’s Disease 2020 |
| Framingham raw-audio AUC, normal cognition vs dementia | 0.805 | Xue et al., Alzheimer’s Research and Therapy 2021 |
| Framingham language model AUC, normal vs dementia and normal vs MCI | 92.6% and 74.4% | Amini et al., Alzheimer’s and Dementia 2023 |
| Predicting MCI-to-Alzheimer’s progression within 6 years | AUC 78.5% (166 participants) | Amini et al., Alzheimer’s and Dementia 2024 |
| Speech AUC for amyloid positivity in cognitively unimpaired adults | 0.71 | Fristed et al., Brain Communications 2022 |
| Acoustic model vs neuropsychological tests for amyloid status in MCI | AUC 0.79 vs 0.66 | Garcia-Gutierrez et al., Frontiers in Neuroscience 2023 |
| Average MCI detection rate in US primary care | 0.08 of expected cases | Liu et al., Journal of Prevention of Alzheimer’s Disease 2024 |
| Medicare beneficiaries meeting MCI criteria with a recorded diagnosis | 10.6% | Couch et al., Alzheimer’s and Dementia DADM 2026 |
| Americans 65 and older living with Alzheimer’s dementia | 7.4 million (2026) | Alzheimer’s Association 2026 |
| Recommended floor for a brief detection tool, sensitivity and specificity | 80% and 85% | Global CEO Initiative on Alzheimer’s Disease 2025 |
| First FDA-cleared Alzheimer’s blood test, positive and negative agreement | 91.7% and 97.3% | FDA, May 16, 2025 |
Methodology and Sources
We ran targeted searches and pulled figures only from primary documents read in full or at abstract level: peer-reviewed papers on PubMed Central and Europe PMC, the ADReSS challenge papers on arXiv, the DementiaBank repository pages, agency pages and the Alzheimer’s Association. No figure comes from a secondary blog. Where a paper reports a number inconsistently, we used the abstract figure and noted the conflict below.
- Jafari Z, Andrew MK, Rockwood KJ. Diagnostic utility of speech-based biomarkers in mild cognitive impairment: a systematic review and meta-analysis. Age and Ageing, 2025.
- Cohen AS et al. Speech pause and speech rate for evaluating Alzheimer’s and mild cognitive impairment: a meta-analysis. Journal of the International Neuropsychological Society, 2026 (DOI 10.1017/s1355617725101677, abstract read via Europe PMC).
- Saeedi S et al. Acoustic Speech Analysis in Alzheimer’s Disease: A Systematic Review and Meta-Analysis. Journal of Prevention of Alzheimer’s Disease, 2024.
- Luz S et al. Alzheimer’s Dementia Recognition through Spontaneous Speech: The ADReSS Challenge. Interspeech, 2020.
- Luz S et al. Detecting cognitive decline using speech only: The ADReSSo Challenge. Interspeech, 2021.
- Luz S et al. An Overview of the ADReSS-M Signal Processing Grand Challenge on Multilingual Alzheimer’s Dementia Recognition Through Spontaneous Speech. IEEE Open Journal of Signal Processing, 2024.
- TalkBank DementiaBank home, corpus index and Pitt corpus page, accessed October 2026.
- de la Fuente Garcia S, Ritchie CW, Luz S. Artificial Intelligence, Speech, and Language Processing Approaches to Monitoring Alzheimer’s Disease: A Systematic Review. Journal of Alzheimer’s Disease, 2020.
- Xue C et al. Detection of dementia on voice recordings using deep learning: a Framingham Heart Study. Alzheimer’s Research and Therapy, 2021.
- Amini S et al. Automated detection of mild cognitive impairment and dementia from voice recordings: a natural language processing approach. Alzheimer’s and Dementia, 2023 (DOI 10.1002/alz.12721, abstract read via Europe PMC).
- Amini S et al. Prediction of Alzheimer’s disease progression within 6 years using speech. Alzheimer’s and Dementia, 2024.
- Park S et al. Linguistic features from paragraph recall are markers of cognitive impairment. Journal of the International Neuropsychological Society, 2026.
- Fristed E et al. AMYPRED study of speech and artificial intelligence screening for early Alzheimer’s disease and amyloid beta positivity. Brain Communications, 2022.
- Garcia-Gutierrez F et al. Acoustic speech parameters and amyloid status in individuals with mild cognitive impairment. Frontiers in Neuroscience, 2023.
- van den Berg RL et al. Digital remote assessment of speech acoustics in cognitively unimpaired adults. Alzheimer’s Research and Therapy, 2024.
- Hale MR et al. Linking silent pause locations in speech to amyloid and tau deposition in cognitively unimpaired individuals. Journal of Alzheimer’s Disease, 2025 (DOI 10.1177/13872877251371720, abstract read via Europe PMC).
- Garcia AM. Gauging speech and language markers of dementia against pathological and neuroimaging data. Alzheimer’s and Dementia, 2025 (conference abstract).
- WHO Dementia Fact Sheet, updated July 3, 2026.
- Alzheimer’s Association, 2026 Alzheimer’s Disease Facts and Figures.
- Liu Y et al. Detection Rates of Mild Cognitive Impairment in Primary Care for the United States Medicare Population. Journal of Prevention of Alzheimer’s Disease, 2024 (DOI 10.14283/jpad.2023.131, abstract read via Europe PMC).
- Couch E et al. Prevalence and predictors of diagnosed mild cognitive impairment among Medicare beneficiaries. Alzheimer’s and Dementia: Diagnosis, Assessment and Disease Monitoring, 2026.
- Thompson LI et al., Global CEO Initiative on Alzheimer’s Disease. Acceptable standards for clinic-based digital cognitive assessments. Alzheimer’s and Dementia, 2025.
- 21 CFR 882.1470, Computerized cognitive assessment aid, eCFR.
- US FDA. FDA Clears First Blood Test Used in Diagnosing Alzheimer’s Disease, May 16, 2025, and Breakthrough Devices Program statistics as of June 30, 2026.
- Data watch: The Age and Ageing meta-analysis is internally inconsistent in places: its abstract gives 21 studies and a 78.0% pooled AUC, while the body counts 23 studies and the discussion cites 76.0%; it also reports I-squared = 0.00% for every metric under a fixed-effects model despite describing highly varied methods, so its confidence intervals may look tighter than the underlying evidence. The ADReSS-M overview gives 225 English training speakers in its data table but 237 training samples in its baseline section. The 2024 Framingham progression paper reports 78.5% accuracy in its abstract but 78.8% in its results table. The Long Life Family Study paper gives 14% impaired in its methods but 16% in its table (112 of 710 is 15.8%). The study-quality audit (de la Fuente Garcia et al.) reports that 77% of its 51 studies did not state data availability but also counts 39 such studies (39 of 51 is 76.5%), so we left that figure out of the tables. That audit and the Framingham raw-audio study date from 2020 and 2021 and are flagged as most recent available data. Burden figures in this roundup cover the United States and global totals only. Several pathology-anchored studies were small, single-site or industry-funded, and none of the speech tools here has been validated as a stand-alone diagnostic.
Last updated: October 24, 2026. We update this roundup quarterly, and the next refresh is expected when the Alzheimer’s Association publishes its 2027 Alzheimer’s Disease Facts and Figures report.