Estimating MMSE and MoCA from speech in clinical trials

Yes, MMSE and MoCA can be estimated from speech. A model trained on recorded speech and rated scores reads a short task, such as describing a picture, and returns an estimate on the scale. It is checked on people it never saw in training and sits beside the rater’s score, not in its place.

Answer page for trial sponsors, clinical operations and digital endpoint teams · Updated November 2026

How it works

From a speech task to a score on the scale

MMSE and MoCA are scored from 0 to 30. A speech model is a regression onto that range, so its estimate is a value in scale points, set beside the rater’s whole‑number score.

  1. A short speech task

    The participant describes a picture, reads a passage, tells a memory or names words in a category. Picture description of the Cookie Theft scene is a widely studied task.

  2. Voice and language features

    Timing and pauses, pitch and loudness, word choice, word‑finding and how much of the scene is described, taken from the audio and its transcript, often with deep‑learning representations beside them.

  3. Regression onto the scale

    A model learns how those features relate to MMSE or MoCA scores rated in the same session, across the whole range from healthy to dementia.

  4. Checked on new people

    The model is scored on participants it never saw in training, in the language and setting where it will be used, and reported in scale points.

Public research

What has been published

Shared challenges made MMSE estimation from speech a common research task. Separate studies have estimated MoCA from other tasks, and reviews set out how such measures should be evaluated.

Shared challenges on MMSE estimation

  • Luz S, Haider F, de la Fuente S, Fromm D, MacWhinney B. Alzheimer’s Dementia Recognition Through Spontaneous Speech: The ADReSS Challenge. Proc. Interspeech 2020, 2172‑2176. doi:10.21437/Interspeech.2020-2571

    Set two tasks on recorded Cookie Theft picture descriptions: telling dementia from healthy speech, and estimating each speaker’s MMSE. Groups were balanced by age and gender, and the audio was cleaned of noise and levelled to reduce recording differences.

  • Luz S, Haider F, de la Fuente S, Fromm D, MacWhinney B. Detecting Cognitive Decline Using Speech Only: The ADReSSo Challenge. Proc. Interspeech 2021, 3780‑3784. doi:10.21437/Interspeech.2021-1220

    Moved to speech only, with no manual transcripts provided: dementia detection, MMSE estimation and prediction of later cognitive decline, from features extracted directly from the recordings.

  • Luz S, Haider F, Fromm D, Lazarou I, Kompatsiaris I, MacWhinney B. 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;5:738‑749. doi:10.1109/OJSP.2024.3378595

    Trained on English picture descriptions and tested on Greek descriptions of a different picture, for dementia detection and MMSE estimation. Some features carried across languages, and the organizers call for further research on transfer between languages.

  • Balagopalan A, Eyre B, Rudzicz F, Novikova J. To BERT or not to BERT: Comparing Speech and Language‑Based Approaches for Alzheimer’s Disease Detection. Proc. Interspeech 2020, 2167‑2171. doi:10.21437/Interspeech.2020-2557

    On the ADReSS data, compared hand‑crafted acoustic and linguistic features with fine‑tuned language models, and fitted regression models for the cognitive score.

Studies estimating MoCA

  • Wisler AA, Fletcher AR, McAuliffe MJ. Predicting Montreal Cognitive Assessment Scores From Measures of Speech and Language. Journal of Speech, Language, and Hearing Research. 2020;63(6):1752‑1761. doi:10.1044/2020_JSLHR-19-00183

    Older adults described an early childhood memory. Lexical variation and speaking rate related to MoCA scores, and the authors note that a free narrative carries only part of the information in the score, a reason task choice matters.

  • Cay G, Pfeifer VA, Lee M, et al. Harnessing Speech‑Derived Digital Biomarkers to Detect and Quantify Cognitive Decline Severity in Older Adults. Gerontology. 2024;70(4):429‑438. doi:10.1159/000536250

    Participants read a standard passage aloud into a wearable microphone. Timing, pause and similarity features related to MoCA, and a random forest model estimated MoCA scores from them.

Reviews, frameworks and the scales

  • 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;78(4):1547‑1574. doi:10.3233/JAD-200888

    A systematic review of speech and language approaches to Alzheimer’s disease; it points to poor standardisation and the gap between study aims and clinical use.

  • Robin J, Harrison JE, Kaufman LD, Rudzicz F, Simpson W, Yancheva M. Evaluation of Speech‑Based Digital Biomarkers: Review and Recommendations. Digital Biomarkers. 2020;4(3):99‑108. doi:10.1159/000510820

    Applies the V3 framework to speech: recording quality and hardware, test‑retest reliability and demographic variation, then correspondence with diagnosis, progression or treatment response.

  • Folstein MF, Folstein SE, McHugh PR. “Mini‑mental state”. Journal of Psychiatric Research. 1975;12(3):189‑198. doi:10.1016/0022-3956(75)90026-6

    The original MMSE, a bedside test of orientation, attention, memory, language and visuospatial skills.

  • Nasreddine ZS, Phillips NA, Bédirian V, et al. The Montreal Cognitive Assessment, MoCA: A Brief Screening Tool For Mild Cognitive Impairment. Journal of the American Geriatrics Society. 2005;53(4):695‑699. doi:10.1111/j.1532-5415.2005.53221.x

    The original MoCA, a brief screening test for mild cognitive impairment.

What to check

Six questions before a sponsor relies on the estimate

A speech estimate of MMSE or MoCA is only as useful as the evidence for your population, your language and your recording setup. For a wider provider checklist, see choosing a speech biomarker provider.

  1. Was it tested on people the model never saw?

    A model scored on the people it was fitted to looks better than it will in a new trial. Ask for results on held‑out participants, ideally from another cohort, with error in scale points. Ask for that error inside your trial’s own entry range too, since a fit across the whole healthy‑to‑dementia range looks closer than one inside a narrow band.

  2. Is it calibrated in each study language?

    Speech differs by language; in the ADReSS‑M challenge, some features carried from English to Greek. Ask how each study language is calibrated, and on which rated scores.

  3. Which speech task, and is it the same every time?

    Picture description, reading aloud and free narrative draw on different abilities. The same task at every session keeps readings comparable; ask how practice from repeating it is handled.

  4. How is the audio recorded?

    Microphone, room noise and the route, site, app or phone call, all shape the audio. Ask which setups the model was checked on and which quality rules reject a recording.

  5. What happens when phones or their software change?

    Operating system updates, noise suppression and new phones change the audio before analysis. The FDA’s December 2023 guidance on digital health technologies asks sponsors to record updates and compare data before and after. Phone changes and speech endpoints ›

  6. How is it reported beside the rater’s score?

    The estimate should come in scale points, joined to the visit, with the named features behind it. The rated MMSE or MoCA stays the reference; the speech estimate adds readings beside it and does not replace the rater.

Use in trials

Where a speech estimate fits in a trial

MMSE and MoCA are rated at a handful of visits. A short speech task can add readings on the same scale at three points in a study.

Pre‑screening against the entry range

A short recording on the candidate’s own phone or by phone call, read against the MMSE or MoCA range in the entry criteria before the screening visit. The screening visit still decides eligibility.

Readings between visits

Home sessions add estimates between rated visits, on the same scale, so cognition is read more often than the visit schedule allows.

Beside each site rating

At rated visits the estimate sits next to the rater’s score. A large gap flags that visit for review of the recording and of how the test was given. Trial QC ›

Questions

MMSE and MoCA from speech

Can MMSE or MoCA be estimated from speech?

Yes. A model trained on recorded speech and rated scores can estimate MMSE or MoCA from a short speech task such as describing a picture. The ADReSS and ADReSSo challenges made MMSE estimation from speech a shared research task in 2020 and 2021, and peer-reviewed studies have estimated MoCA from narratives and from reading aloud.

Which speech tasks are used to estimate MMSE or MoCA?

Published work uses picture description, often the Cookie Theft picture, reading a standard passage aloud, telling a personal memory, verbal fluency and short memory questions. Each task draws on different abilities, so the same task should be used at every session of a study.

Does a speech estimate replace the MMSE or MoCA in a trial?

No. The rated MMSE or MoCA stays the reference, and the speech estimate sits beside it: before the screening visit to check a candidate against the entry range, between visits, and next to each site rating.

Does MMSE or MoCA estimation from speech work in languages other than English?

It needs calibration in each language. The ADReSS-M challenge trained models on English and tested them on Greek; some features carried across, and the organizers call for further research on transfer between languages. Ask a provider how each study language is calibrated and on which ratings.

Why is a speech estimate of the MMSE or MoCA not a whole number?

MMSE and MoCA are wide-range scales scored from 0 to 30, so the model is a regression and returns a value on that range. It is reported in scale points beside the rater’s whole-number score.

What should a sponsor ask before using a speech estimate of MMSE or MoCA?

Ask whether it was tested on people the model never saw, how each language is calibrated, which speech task is used, how the audio is recorded, what happens when phones or their software change, and how the estimate is reported beside the rater’s score.

Cephalgo

How Cephalgo does this

Cephalgo, a speech biomarker company in Strasbourg, France, estimates MMSE or MoCA from a short speech task, in recordings a trial already holds, at site visits or in home sessions. Each estimate comes with the named speech features behind it, for memory, word‑finding and fluency, and is joined to the visit score. Readings stay stable across phone and software changes, and each study language is calibrated on the study’s own screening and baseline ratings, blind to arm. The model was tested on participants it never saw in training: MMSE on a public challenge test set, MoCA in an external cohort. Alzheimer’s programs › Validation approach › Speech biomarker companies ›

Estimate MMSE or MoCA in your study

Share your scale, your entry range and any audio you hold, and get a fixed‑fee scope in writing.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.