/ 04
Four signals, read honestly
What the evidence actually supports for each one, not what would be convenient to claim.
Each card below separates three different questions: does voice measure this at all, does this predict something that matters, and what did our own pilot find. The validation tier comes from Amplifier's own internal classification, not marketing language. Border colors mark the tier: green for established, rust for investigational, blue for a metric that isn't a classifier at all.
Fatigue
Established sign · in-domain AUC 0.818R&D candidate model, not yet in production. Production runs an earlier, lighter model with its own unpublished accuracy figure.
Validity
0.95 / 0.93
Train / test AUC classifying high vs. low fatigue from pitch, pause, and utterance duration against a validated multiple sclerosis fatigue scale (Nylander et al., 2024, Multiple Sclerosis)
Predicts decline
OR 2.14
Mortality odds ratio for fatigue across 30 studies and 152,711 older adults; physical decline shown to precede hospitalization and death (Knoop et al., 2021, Ageing Research Reviews)
Our pilot
+52%
Rise in mental fatigue during an induced stress phase, tracking self-report at 91% agreement, the highest of any dimension measured
Stress
Established sign · in-domain AUC 0.884Validity base is strongest for acute, lab-induced stress against cortisol. Ambient daily-life validation is thinner in the published literature.
Validity
58.8%
Of variance in peak cortisol rise explained by acoustic features under a standard lab stress protocol, 80% classification accuracy (Oesten et al., 2023, IEEE EMBS BHI)
Predicts decline
OR 2.81
Persistent psychosocial stress and acute myocardial infarction, on par with hypertension in the same cohort (Lanas et al., 2007, Circulation)
Our pilot
+48%
Rise in stress during the induced phase. Pitch variability was the single strongest acoustic predictor of this dimension
Hydration
Sign · in-domain AUC 0.938The highest raw accuracy of the three classifiers and the thinnest literature behind it. Read the flag below before using this one with a technical buyer.
Validity
Mixed
Surface vocal-fold hydration changing jitter and HNR is well established (Mahalingam & Boominathan, 2016, Laryngoscope); a purpose-built systemic-dehydration study found confirmed physiologic dehydration with no detectable acoustic signature (Rodgers et al., 2025, PLOS One)
Predicts decline
OR 3.53
Dehydration on admission and delirium in adults over 65; delirium itself carries a 4.56x mortality odds ratio out to two years (Pendlebury et al., 2015, BMJ Open)
Our pilot
Not measured
Hydration was not one of the seven dimensions tracked in our internal validation study. No internal proof point exists yet for this signal
What we won't claimExisting bedside hydration checks in older adults are documented to catch only about 1 in 5 confirmed cases (Munk et al., 2021, Clin Nutr ESPEN), so the bar this signal has to clear is low. That's a reason to keep developing it, not a reason to claim voice has cleared it yet. We'd frame this as a promising, actively researched signal, not a validated one.
Energy
Extended metric · not a classifier, no AUCBy design, this sits outside the flagged signal pipeline. It is observational wellness data, reported but never used to trigger a flag on its own.
Validity
Adjacent
No study validates voice directly against a standard vitality scale. The closest evidence is psychomotor retardation research in depression, a related but distinct construct (Flint et al., 1992, 1993)
Predicts decline, generally
HR 0.60
Per one standard-deviation increase in a standard vitality scale, for all-cause mortality over 8.6 years, in 6,987 adults. This is vitality as a construct, not yet tied to a voice measurement of it (Andersen et al., 2021, J Psychosomatic Research)
Our pilot
−35%
Drop in energy during the induced stress phase, at 89% self-report agreement, the second-highest of any dimension measured
What we won't claimThe internal pilot number is genuinely good. But structurally this is the softest of the four claims: no classifier, no AUC, no direct external validation against a named instrument. We'd present it as a wellness indicator that tracks closely with self-report, not as clinically grounded as the three signs above it.