Testing Cross-Species Alarm-Call Features as a Cortisol-Grounded Stress Signal in Human Speech

Animals that have never heard each other's calls often still react correctly to them. One proposed explanation is that some acoustic properties carry meaning on their own. Harsh, noisy, irregular voicing seems to signal arousal, and faster pulse rates seem to signal urgency, across many species. People who build speech-based stress detectors ignore this idea. They extract thousands of generic acoustic statistics and fit them to whatever label is available, usually a questionnaire. Questionnaires are biased: people shade answers toward what they think the experimenter wants, toward social norms, or toward a misremembered past.

This proposal joins those two gaps. We pre-specify a small set of acoustic features that animal-communication theory predicts should carry arousal. We fix the expected direction and timing of each feature's relationship to stress before looking at data. We then test whether these features track salivary cortisol, a hormone measured in the body and not reported by the person. The key test is a deliberate manipulation of what participants believe the study is about. If the features track the body rather than the report, self-reported stress should shift under that manipulation while cortisol and the conserved features should not. The result would be a falsifiable, interpretable stress measure, or a clean negative showing that the animal-derived features add nothing beyond standard speech features.

Core Claim

The claim is that a compact, theory-fixed set of "conserved distress primitives" predicts between-person differences in cortisol response at least as well as brute-force feature banks. It also stays stable when self-report is pushed around by demand characteristics, meaning participants change their answers once they guess the study's purpose. Three properties are new together: the features are chosen in advance from cross-species theory, they are validated against an endocrine measurement, and the comparison with self-report is experimental rather than correlational.

The primitives come from three sources in the file. The first is the "featural interpretation" route of Anti‐Babel: Three degrees of interspecies comprehension. That paper argues on theoretical grounds that nonlinear voicing (subharmonics, deterministic chaos, noise-like segments) maps to arousal and pulse rate maps to urgency. It never tests these mappings acoustically, and it never touches human physiology.

The second source is the infant study Rhythm May Be Key to Linking Language and Cognition in Young Infants. It found that temporal-envelope rhythm (AUC 0.9939) and spectral envelope (AUC 0.9955), not pitch (AUC 0.6287), separate sounds that infants treat like speech, including lemur calls. This supplies a prior that envelope and rhythm features are shared across primates. It also supplies a warning against relying on pitch.

The third source is Acoustic correlates of the syllabic rhythm of speech. It showed that global modulation spectra need 8–9 minutes of speech to stabilize. A local envelope trough about 32 ms after syllable onset, by contrast, holds across eight languages. This constrains the design: TSST speech segments are short, so we use local time-domain envelope features rather than global modulation spectra.

The resulting set has about ten features. Each one has a pre-registered sign: for example, more subharmonic and chaotic voicing goes with a higher cortisol rise, and faster local envelope pulsing goes with a higher cortisol rise.

Research Program

The first component is a re-analysis of the FAU-TSST corpus. It contains 43 speakers recorded during the Trier Social Stress Test, a standard lab protocol of public speaking and mental arithmetic, with saliva sampled repeatedly up to 60 minutes after the stressor. Using Speech to Predict Sequentially Measured Cortisol Levels During a Trier Social Stress Test predicted cortisol from 6,373 hand-crafted features. Its best result was a Spearman ρ of 0.421 (RMSE 0.245) at 20 minutes after the stressor. The 88-feature EGEMAPS set reached ρ = 0.449 in mental arithmetic at 10 minutes. That work motivates ours: it shows speech carries endocrine information. It leaves open which acoustic properties do the work, and whether they mean anything beyond the corpus.

One confound matters before anything else. Cortisol follows a stereotyped time course after a stressor, so a model can score well by recognizing which sampling point a recording came from. All our primary analyses are therefore within a single time point, ranking people against each other. We also report the published pooled metrics only for comparability.

The second component is a new data collection, which carries the self-report question. We run about 120 participants through the Trier test, recording speech, salivary cortisol, and self-rated stress at each sampling point. Participants are randomized to two cover stories. One frames the study as measuring "resilience under pressure," which invites participants to under-report stress. The other is neutral. This is a direct manipulation of demand characteristics, the distortion the file documents in clinical trials and surveys.

The prediction is a dissociation. The cover story should shift self-reported stress. It should not shift cortisol, and it should not shift the primitive-based estimate. The question is not "do the features correlate with stress." It is "which measure moves when only the report is biased."

The third component is a causal perturbation check, in the spirit of the "pitting experiments" Anti-Babel says are missing. We resynthesize calm speech from the corpus and add controlled amounts of each primitive, such as injected subharmonics or altered envelope pulse rate, while holding pitch and words fixed. We then check that the trained model's predicted stress moves in the pre-registered direction and roughly linearly with dose. A model that has learned recording or speaker artifacts rather than the primitives will not do this.

As a secondary step, we compute the same primitives on the audible band of the MADUV mouse recordings. MADUV records pups during maternal separation, which is itself a distress context, and reports that the audible band below 20 kHz beats the ultrasonic band. This step asks only whether the primitives are measurable and informative outside humans. We treat it as a plausibility check, not as evidence of homology.

Baselines

The naive reference is a model that knows only the sampling time point and the group's mean cortisol curve. Within a time point it predicts the same value for everyone, so its between-person ρ is zero by construction. Its pooled ρ shows how much of the published numbers is the time-course confound.

A second reference is self-reported stress used directly as the predictor of cortisol. This is what a questionnaire-based study implicitly assumes. Beating it in the biased-cover-story arm, while matching it in the neutral arm, is the result that matters for the self-report direction.

The strongest existing method is support vector regression on COMPARE features, as in the FAU-TSST cortisol study. It is the published best on this exact target at 20 minutes: ρ = 0.421, RMSE 0.245. EGEMAPS with the same regressor is the second baseline, with ρ = 0.449 in mental arithmetic at 10 minutes. We rerun both, plus the primitive set, under identical conditions: leave-one-speaker-out folds, the same per-speaker normalization, the same regressor family, and the same hyperparameter search budget, all nested inside the training folds.

The proposal beats these baselines under either of two outcomes. One is matching their within-time-point ρ with roughly a hundredth of the features. The other is exceeding them on the new cohort, where the primitives are applied unchanged while the baselines must be refit.

Two ablations isolate the key ingredient. The first removes the nonlinear-voicing and local-envelope primitives and keeps only pitch and energy. The second adds those primitives to EGEMAPS and asks whether they explain variance EGEMAPS misses. If the primitives add nothing to EGEMAPS, the specific animal-derived content of the claim fails. That is a publishable result in its own right.

Why not just…?

Why not just use EGEMAPS, which is already a small, expert-chosen set? This is the real threat. EGEMAPS already contains jitter, shimmer and harmonics-to-noise ratio, which partly capture nonlinear voicing, and it reached the best published single result on FAU-TSST. The proposal does not survive by argument here. It survives only if the incremental-variance ablation shows that subharmonic, chaos and local-trough measures improve within-time-point prediction beyond EGEMAPS. It also needs the perturbation test to show that the model responds in the predicted direction. If neither holds, we will report that the cross-species framing adds interpretation but no measurement value.

Why not just fine-tune a pretrained speech or bioacoustic encoder? With 43 speakers, a large encoder mostly memorizes speakers. The FAU-TSST study found deep spectrum features consistently below COMPARE. AVES found that 12-layer models beat 24-layer ones on downstream bioacoustic tasks because of overfitting. We still include a frozen AVES embedding with a linear probe under the same folds, as an empirical check. A pretrained encoder also cannot give the pre-registered directional predictions the dissociation test needs.

Why not just use better questionnaires or momentary mood prompts? In the intern micro-randomized trial, the mobile intervention study found that the measured mood response itself depended on prior mood. Demand effects are documented throughout the file. More importantly, the cover-story manipulation measures directly how much self-report moves under bias. If self-report turns out to be stable across arms, this alternative wins, and we will say so.

Why not just analyze transcripts? On DAIC-WOZ, a depression-interview benchmark, the prompt-validity study showed that a text model reaches F1 = 0.88 from the virtual interviewer's prompts alone. Participant-only text reaches only 0.64. Transcripts invite exactly the shortcuts we want to avoid. As a side check, we compute primitives on participant-only DAIC-WOZ audio and compare them against audio that includes interviewer turns. No reported acoustic result was found on Lacuna for DAIC-WOZ, so no prior number exists for that comparison.

Decisive Experiments

Three results settle the proposal. On FAU-TSST, within each time point and under leave-one-speaker-out evaluation, the primitive set must reach a ρ statistically indistinguishable from or above COMPARE and EGEMAPS refit under the same protocol. It must also add significant variance over EGEMAPS.

In the new cohort, the cover story must produce a significant group difference in self-reported stress. It must produce no detectable difference in cortisol rise or in the primitive-based estimate. The primitive estimate's correlation with cortisol must hold in both arms, while self-report's correlation degrades in the biased arm.

In resynthesis, predicted stress must change monotonically and in the pre-registered direction as each primitive is added.

Failing the first result, while passing the second through EGEMAPS too, would mean ordinary paralinguistic features already resist self-report bias. That would still be useful, but it would not support the cross-species claim. Failing the third would mean the model's apparent success rests on artifacts.