Abstract
Phonation, or the vibration of the vocal folds, is the primary source of vocalization in the production of voiced sounds by humans. It is a complex bio-mechanical process that is highly sensitive to changes in the speaker's respiratory parameters. Since most symptomatic cases of COVID-19 present with moderate to severe impairment of respiratory functions, we hypothesize that signatures of COVID-19 may be observable by examining the vibrations of the vocal folds. Our goal is to validate this hypothesis, and to quantitatively characterize the changes observed to enable the detection of COVID-19 from voice. For this, we use a dynamical system model for the oscillation of the vocal folds, and solve it using our recently developed ADLES algorithm to yield vocal fold oscillation patterns directly from recorded speech. Experimental results on a clinically curated dataset of COVID-19 positive and negative subjects reveal characteristic patterns of vocal fold oscillations that are correlated with COVID-19. We show that these are prominent and discriminative enough that even simple classifiers such as logistic regression yields high detection accuracies using just the recordings of isolated extended vowels.
Introduction
While many researchers have attempted to detect COVID-19 from voice and cough recordings, most rely on "black-box" deep learning models. These models often struggle with small datasets and are prone to overfitting—learning speaker-specific quirks rather than actual disease markers. This research takes a physics-first approach, hypothesizing that COVID-19, as a respiratory illness that impairs lung function and airway musculature, leaves a measurable "signature" in vocal fold vibrations.
Phonation relies on an intricate balance of aerodynamic forces across the vocal folds. Because direct clinical imaging of vocal folds is difficult at scale, this study utilizes mathematical modeling to "see" these vibrations through recorded audio. By reconstructing vocal fold motion from simple voice recordings, the study found that COVID-19 significantly disrupts the symmetry and synchronization of speech production, allowing for detection with high accuracy.
Biomechanical Modeling and Parameter Estimation
The researchers utilize a 1-mass asymmetric body-cover model, a mathematical framework that describes the horizontal displacements of the left and right vocal folds using coupled non-linear differential equations.
To bridge the gap between raw audio and these physical equations, they employ the ADLES (Acoustic-to-Dynamic Lossless Estimation of Speech) algorithm. This approach jointly solves two problems:
- The Forward Problem: Predicting vocal fold motion based on physical parameters such as mass, spring tension, and asymmetry.
- The Inverse Problem: Estimating those physical parameters by comparing the model’s predicted glottal flow to the actual flow extracted from a voice recording via inverse filtering.
By minimizing the "residual energy" (the difference between the model and the real-world signal), the algorithm extracts features like the coupling coefficient ($\alpha$), the asymmetry coefficient ($\Delta$), and the residual error ($R$).
Experimental Setup
The study used a clinically curated dataset from Merlin Inc. (Chile), consisting of 512 individuals tested for COVID-19. A selected subset of 19 individuals (10 female, 9 male) was used for this analysis, all recorded within 7 days of their medical test. The group included 9 COVID-positive and 10 COVID-negative subjects. Audio data consisted of recordings of isolated extended vowels (/a/, /i/, and /u/) sampled at 8 kHz.
Analysis of Vocal Fold Dynamics
A key finding of the study is the disruption of phase space trajectories in COVID-19 patients. In healthy individuals, vocal fold trajectories in phase space (displacement vs. velocity) appear as "limit cycles" or "slim toroids," indicating highly synchronized and symmetrical motion.
In COVID-19 patients, these trajectories become complex and "messy," signaling a loss of synchronization and a reduced range of motion. Interestingly, the most discriminative feature was the residual $R$. This suggests that standard mathematical models for healthy speech cannot account for the "anomalous" vibrations caused by the infection; the model's failure to fit the data effectively becomes the biomarker.
Furthermore, the vowel /i/ (a high front vowel) was significantly more effective for detection than /a/ or /u/. This is likely because producing /i/ requires reaching specific high-frequency energy peaks that are particularly compromised by respiratory impairment.
Results and Discussion
The researchers tested several classifiers, including Logistic Regression (LR), Support Vector Machines (SVM), and Random Forests (RF). Most classifiers achieved an area under the curve (ROC-AUC) of approximately 0.80 across all vowels. However, Logistic Regression using only the vowel /i/ achieved a remarkable AUC of 0.912. All results were found to be highly significant (p-values < $1e^{-5}$), suggesting that these biomechanical features are robust indicators of the disease.
Despite these promising results, the authors note two major caveats. First, the study does not yet prove that these vocal anomalies are exclusive to COVID-19; other respiratory conditions like the flu might produce similar disruptions. Second, the method is likely only effective for symptomatic patients whose respiratory functions are already impaired, meaning it may not serve as a tool for detecting asymptomatic carriers.
Conclusion
This research demonstrates that COVID-19 leaves a distinct biomechanical signature on vocal fold oscillations. By moving beyond black-box models toward physics-based estimation, the study provides a more interpretable and potentially robust method for disease detection. Future work will focus on using more sophisticated dynamical models to capture pathological oscillations and expanding the dataset to ensure diagnostic specificity against other respiratory illnesses.
Summary updated