Abstract
COVID-19 often impairs respiratory functions, leading to asynchronous and asymmetrical vocal fold oscillations during phonation. This study proposes a method to detect COVID-19 by analyzing the differential dynamics of the glottal flow waveform (GFW). By comparing a GFW estimated from recorded speech against a GFW derived from a physical model of healthy phonation, we identify anomalies indicative of the disease. A CNN-based "Sandwiched" 2-step attention model (S2AP) is employed to locate these anomalies in time-feature space. Using a clinically curated dataset, the method achieves an ROC-AUC of up to 0.90, demonstrating the potential of voice-based COVID-19 screening.
Introduction
The pathogenesis of COVID-19 frequently involves respiratory distress, which manifests in the mechanics of voice production. Clinical observations suggest that infected individuals exhibit irregular vocal fold vibrations. However, direct measurement of these oscillations is invasive and impractical for large-scale screening. A significant hurdle in acoustic analysis is the lack of "baseline" recordings; without knowing a person's healthy voice, it is difficult to distinguish pathological changes from natural vocal characteristics.
This research addresses this by using a physical model of phonation as a universal "healthy proxy." By measuring the divergence between a patient's actual glottal airflow and the predicted airflow of a healthy physical model, the system can detect respiratory anomalies without requiring prior recordings of the individual.
Methodology
The proposed detection pipeline consists of three primary stages: glottal flow estimation, residual analysis, and classification via a specialized neural network.
Glottal Flow Estimation and Residual Analysis
The system extracts the Glottal Flow Waveform (GFW) using two distinct approaches:
- Inverse Filtering ($GFW_{filter}$): Estimates the raw airflow from the recorded audio signal using Iterative Adaptive Inverse Filtering (IAIF).
- Physical Modeling ($GFW_{model}$): Uses the ADLES algorithm and a 1-mass asymmetric body cover model to simulate healthy vocal fold oscillations.
In a healthy subject, these two waveforms should closely align. In COVID-19 patients, the physical model—which assumes normalcy—fails to track the irregular, asynchronous airflow, resulting in a "residual" error signal that contains discriminative features.
Sandwiched 2-step Attention Pooling (S2AP)
To analyze the residual signal, the authors developed the Sandwiched 2-step Attention Pooling (S2AP) architecture. While standard 2-step attention pools information across frequency and then time, the "Sandwiched" version inserts an additional feature detector between these steps. This allows the model to better capture specific temporal anomalies, such as phase shifts or sudden glitches in vibration, which are characteristic of COVID-19.
Results and Discussion
The model was evaluated using a clinically curated dataset of recordings (vowels /a/, /i/, /u/) from 19 subjects, collected within seven days of medical testing.
Vowel Sensitivity
A key finding of the study is that the model's performance varies significantly depending on the vowel phonated. The system showed higher sensitivity to high vowels like /i/ and /u/ compared to the low vowel /a/.
As shown in the table above, the combination of /i/ + /u/ yielded the highest performance, achieving an ROC-AUC of 0.900. The authors hypothesize that the vocal fold configurations required for these vowels may be more susceptible to the subtle physical impairments caused by COVID-19.
Architecture Performance
The transition from standard attention mechanisms to the S2AP architecture provided a substantial boost in accuracy. Without additional CNN layers, S2AP improved the ROC-AUC from 0.66 to 0.79. The best-performing configuration, utilizing a CNN (2,5,64) with S2AP, reached an ROC-AUC of 0.8520 on the general task.
Conclusion
By leveraging a physical model as a "healthy proxy," this method bypasses the need for longitudinal patient data. The S2AP architecture provides both high diagnostic accuracy and a degree of interpretability by highlighting specific moments of desynchronization in the glottal flow. Future work will focus on expanding the dataset and testing the model's ability to differentiate COVID-19 from other respiratory illnesses.
Summary updated