Speech Health Signals and High-Stakes Boundaries

Audio screening for infection is the cleanest form of the promise. Someone coughs or speaks into a phone, a classifier returns a probability, and nobody has to be swabbed. The promise has been tested at scale. The UK Health Security Agency assembled audio from 67,842 individuals with linked metadata, 23,514 of them PCR-positive for SARS-CoV-2. A team led by Harry Coppock trained classifiers on it. Unadjusted, those classifiers separated infected from uninfected audio with an ROC-AUC of 0.846 [0.838, 0.854]. That is the number a product would be sold on.

Then the same team matched participants on measured confounders. The classifiers scored 0.619 [0.594, 0.644]. Their abstract puts it flatly: “after matching on measured confounders, such as age, gender, and self reported symptoms, our classifiers performance is much weaker”. No recording had changed. What had been removed was the company the illness kept: age, gender, self-reported symptoms, and the two national recruitment routes participants had arrived through. One route carried mostly positives. The other carried mostly negatives. The model could learn who had been invited rather than who was ill.

That is the difficulty this lesson keeps returning to. A striking figure from voice is cheap to produce and hard to argue with. So the argument has to be made against the figure rather than with it. Here it was made by the team that produced the figure, on the largest dataset of its kind.

Reading in that order means knowing what the first number was ever a claim about. Speech can contain signals associated with respiratory, neurological, motor, cognitive, or affective conditions. An ROC-AUC of 0.846 [0.838, 0.854] is consistent with that. So is its collapse to 0.619 [0.594, 0.644]. Association in a selected dataset is not diagnosis. Nor prognosis, treatment benefit, or clinically useful screening. The gap between those words is not a matter of house style. The US Food and Drug Administration and the National Institutes of Health keep them as separate defined terms, which the next section quotes.

Which of them is being claimed decides everything downstream. High-stakes health use requires a clearly defined intended use, a reference standard, a prospective workflow, external validation, subgroup evidence, privacy protection, and a plan for false reassurance and false alarm. Those requirements are not the same for a claim about one time point as for a claim about a trajectory or about a clinic. A team chooses between the three before any architecture.

The three are worth separating precisely. Each one has already been tested by somebody, and the tests are not interchangeable.

Cross-sectional association means a score differs between groups at one observed time. That is exactly what the UK study measured on 67,842 individuals. It is the smallest of the three, and it survived confounder matching only as far as 0.619 [0.594, 0.644].

Longitudinal monitoring is a claim about change within one person. It stands or falls on whether the score is stable enough to tell a real change from noise. That has been measured directly. A 2021 cohort study computed intraclass correlation coefficients across gait, balance, voice and tapping tasks in the m-Power data set for Parkinson disease, recorded on smartphones with nobody supervising. Its finding: “Among the features differing between PD and HC, only a few tapping and voice features had good to excellent test-retest reliabilities and medium to large effect sizes. All other features performed poorly in this respect.” — Sahandi Far and colleagues, Journal of Medical Internet Research, 2021. Separating patients from controls is one property of a feature. Staying steady within one person is another. Only a few features had both.

Clinical decision support is a claim that using the score improves what gets decided, or what happens to the patient. Its boundary is statutory rather than editorial. The Federal Food, Drug, and Cosmetic Act takes certain decision-support software outside the definition of a device. But only where the software is “enabling such health care professional to independently review the basis for such recommendations that such software presents so that it is not the intent that such health care professional rely primarily on any of such recommendations to make a clinical diagnosis or treatment decision regarding an individual patient.” — 21 U.S.C. 360j(o)(1)(E)(iii). The same subsection expressly does not exclude software intended to acquire, process or analyse a signal from a signal acquisition system. That is what a voice classifier is. A speech model does not escape device regulation by adding a clinician to the loop.

No two of the three are evidenced the same way. A strong result on the first says nothing at all about the other two.

Four words carry that distinction. Each implies different evidence, a different unit, and a different owner of the decision. Two of them are not the lesson's to define. The FDA and the NIH have jointly maintained a glossary of them since 2016, the BEST Resource — Biomarkers, EndpointS, and other Tools. Any claim made about health from speech has to be measured against its wording.

The discipline is constant; the height of the bar is not. Respiratory monitoring, neurological research, telehealth and clinical documentation set four different bars for the same acoustic signal, and each is answered in its own unit.

Bridge2AI-Voice shows what collection at the top bar looks like. Version 1.0 of the NIH consortium's voice dataset reached PhysioNet in 2025: 12,523 voice-derived recordings from 306 participants across five North American sites. Participants were enrolled by membership of five predetermined groups — respiratory disorders, voice disorders, neurological disorders, mood disorders, and pediatric. The University of South Florida's institutional review board approved collection and sharing, and the release notes also record submission to the University of Toronto's research ethics board.

Two of the release's choices are the lesson. It carries adult-cohort data only; its own data description states that “only data from the adult cohort is available”, with the pediatric data published separately. And it does not distribute the voices at all: “In this release, audio waveforms were omitted, and only spectrograph data and other derived features are made available.” The project has since reached v3.1.0, in 2026, and still contains no raw audio. Access now runs through Synapse with institutional sign-off. Consent and privacy, at this bar, are structural decisions about what leaves the building. Not a paragraph in a policy.

Whichever bar applies, the report has the same shape. It has to carry two kinds of number at once. External and prospective performance by intended-use population can look strong while clinical workflow impact, harm, and patient-reported outcomes point the other way. That contradiction is the finding, not a nuisance to be averaged away.

For a monitoring claim specifically, the number to demand is a reliability coefficient, not a discrimination score. In the m-Power smartphone data, of the features that did differ between groups, only a few tapping and voice features cleared both test-retest reliability and effect size. “All other features performed poorly in this respect.” A monitoring product built on a feature whose intraclass correlation coefficient was never reported is claiming a stability that nobody measured.

The same evidence should also show when a score is tracking the device, the site or the invitation route instead of the patient. That is what the 0.846 [0.838, 0.854] turned out to be riding on. It is also the point at which the system abstains or falls back rather than answering.

An audit is best run as a prosecution. The charge another team should be able to bring against a vocal biomarker claim is that site, device or recruitment artifacts are standing in for disease. Build the audit to meet that charge.

Start by writing the clinical question down. Then write what it quietly assumes: which population, which recording task, which reference standard, which decision. Then find one counterexample to those assumptions — a recruitment route, a device, or a medication that would move the score the same way with no disease behind it. The UK study supplies the template. Two national invitation paths, one reaching people just told they were infected and one sampling a mostly uninfected population, were worth the distance between 0.846 [0.838, 0.854] and 0.619 [0.594, 0.644]. Ask what the equivalent path is here. Then ask for the matched figure, not the raw one.

Then test clinical utility, which asks something different again. Not whether the groups separate, but whether the decision made about the patient improves when the score is used — “a net improvement in health outcome”, in the FDA-NIH wording — and what evidence would show it.

Finish on the limits. If the claim is that a clinician stays in charge, the statute already says what that has to mean. The software must be “enabling such health care professional to independently review the basis for such recommendations”, which a bare probability with no basis attached does not do. Note too that a classifier consuming audio is software that acquires and analyses a signal from a signal acquisition system. The carve-out may not be available at all.