Built on the largest voice
biomarker dataset in the world.

The training corpus was assembled meta-analytically, not scraped. Each of the 35 contributing studies was harmonized to a common recording protocol (sampling rate, channel geometry, prompt taxonomy) and a common consent framework before pooling. Sources span Bridge2AI-Voice (NIH), academic partner cohorts, the University of Florida Perceptual Voice Database, the Saarbrücken Voice Database, and 31 other IRB-registered studies spanning 23 years of respiratory and phonatory research.

No single site contributes more than 18% of the pooled data, which limits site-specific overfitting and gives the reference distribution genuine geographic and demographic breadth.

Before any classification work, the corpus was used to publish normative acoustic and phonatory reference ranges, the healthy distributions of jitter, shimmer, HNR, cepstral peak prominence, formant dispersion, and 40+ additional features, stratified by age, sex, and smoking status. The reference paper appeared in the Journal of Voice (2022) and is what every downstream Rezpio biomarker is compared against.

This matters because it makes Rezpio's "your voice is deviating from your baseline" claim measurable against published norms, not against a private, opaque comparator.

Split-half reproducibility is a test where the recording data is randomly split into two halves, each half is analyzed independently, and the two resulting feature sets are compared. A perfectly reliable measurement returns r = 1.0. Rezpio's features return r = 0.9997 across independent partitions of 12,400 recordings.

In plain language: the measurement is extremely consistent when the same data is split and compared against itself. That is a strong signal of measurement reliability, it means the numbers Rezpio reports on a Tuesday would look almost identical if the pipeline were re-run on a different partition on Wednesday. It is a floor beneath every clinical claim on this page.

The Frontiers in Digital Health preprint reports a mean area under the ROC curve (AUC) of 0.899 across 46 voice biomarkers, evaluated for discriminating stable from decompensating respiratory states.

In plain language: an AUC of 0.5 is chance, an AUC of 1.0 is perfect, and 0.899 sits firmly in the range clinicians recognize as strong discriminative performance, comparable to established triage instruments. It is not a diagnostic verdict; it is evidence that the signal is real, separable, and worth acting on.