Voice-Based COVID-19 Detection: A Physiologically-Grounded Biomarker Approach and the Limits of Transferring the Widmark Alpha Constant - An Analysis of the Coswara-Data Respiratory Sound Corpus (IISc Bangalore)

Published in VoxMedica Review — Inaugural Special Issue: The Vocal Biomarker Ecosystem (Vol. 1, No. 1, July 2026) Abstract This study adapts the methodology developed in a prior voice-alcohol study (BAS Alcohol Language Corpus, "Promil Avcısı" v5) — physiologically interpretable acoustic biomarkers, leakage-free evaluation, and a sex/age-based "alpha constant" (Widmark r) covariate — to the Coswara-Data COVID-19 respiratory sound corpus (Sharma et al., 2020; Bhattacharya et al., 2023). For 2,746 participants, nine recordings each (deep/shallow breathing, heavy/shallow cough, normal/fast counting, three sustained vowels) were extracted from the full 34 GB corpus; biomarkers were computed successfully for 2,715 participants (98.9%) using a sound-type-specific feature set (phonation: F0, jitter, shimmer, HNR, formants; speech: rate/pause; breathing: envelope-periodicity and cycle rate; cough: spectral/MFCC). Critical framing difference from the alcohol study: Coswara contains no blood or serum measurement of any kind. The Widmark r-factor is therefore carried over as an explicitly speculative covariate with no direct pharmacokinetic grounding in this context, and a self-reported symptom-count score substitutes for blood alcohol concentration as a continuous "systemic load" proxy. Findings: (a) numerous biomarkers differ significantly between healthy and COVID-positive participants and remain significant after adjusting for age, sex, and smoking status (p<0.001 for most); (b) the single largest effect size in the entire study, and one with no counterpart in the alcohol work, is reduced breathing-envelope periodicity in COVID-positive participants (Cohen's d=-0.565); (c) the symptom-load×r interaction is significant for only a handful of biomarkers (F0, formant dispersion, speech rate) and the correlation-strengthening effect of r-normalization is small and inconsistent (+0.00 to +0.03, versus a consistent +0.03 to +0.06 across nearly all biomarkers in the alcohol study) — indicating that the alpha constant does not carry the same explanatory weight here; (d) binary classification (healthy vs. COVID-positive) reaches AUC=0.856±0.021 (Gradient Boosting) including demographic covariates, and AUC=0.853±0.023 using acoustic features alone — both substantially stronger than the alcohol study's AUC=0.710.