Improving respiratory disease detection through SSL-enhanced acoustic analysis and exercise-rest measurements

Vera-López Á.; Tilves-Santiago D.; Ramírez-Sánchez J.M.; Docío-Fernández L.; García-Mateo C.; Bustillo-Casado M.; García-Caballero A.A.

doi:10.3389/fmed.2026.1864436

Research / Voice as a clinical signal / [1]

A six-minute walk makes the voice a better diagnostic signal

Improving respiratory disease detection through SSL-enhanced acoustic analysis and exercise-rest measurements. Frontiers in Medicine, 2026.

In one sentence

Recording a patient's cough and voice after light exercise, and describing it with self-supervised speech models, separates post-COVID patients from healthy people with an F1-score of 87.7%.

154participants

2recording moments

3SSL models

87.7%best F1

The problem

Voice is cheap to record and carries information about the lungs and the vocal tract. The changes caused by respiratory conditions are subtle, though, and recordings taken at rest often miss them.

What we did

We recorded 154 participants producing a sustained /a/ and voluntary coughs, once at rest and once after a physiological stress protocol: a six-minute walk and a one-minute sit-to-stand test. Post-Acute Sequelae of SARS-CoV-2 (PASC, or Long COVID) served as the case study.

Each recording was described two ways: with classic acoustic biomarkers, and with embeddings from three self-supervised speech models, wav2vec 2.0, WavLM and HuBERT. A logistic regression classified each participant as PASC or healthy, and a late-fusion step combined the vowel and cough decisions.

Rest / post-exercise audio→ Acoustic + SSL features→ Logistic regression→ Late fusion

What we found

Physical exertion improved classification and reduced variability in every task. Fusing acoustic features with WavLM and wav2vec 2.0 embeddings reached F1-scores of 82.2% for vowels and 80.8% for coughs after exercise, and combining both tasks reached 87.7%.

My contribution

I contributed the acoustic and self-supervised feature pipeline and the experimental design, building on the 2023 study.

Why it matters

A standard walk test and a microphone could become a fast, non-invasive screening step in routine functional assessments for chronic and post-viral conditions. The same recipe, stressing the system before measuring it, applies beyond COVID.

Figure 1. F1-score (%) for PASC vs. healthy classification, post-exercise recordings, stratified 5-fold cross-validation. Screened bars: single tasks. Solid bar: cross-task fusion, the best result.

Cite

Vera-López Á., Tilves-Santiago D., Ramírez-Sánchez J.M., Docío-Fernández L., García-Mateo C., Bustillo-Casado M., García-Caballero A.A. Improving respiratory disease detection through SSL-enhanced acoustic analysis and exercise-rest measurements. Frontiers in Medicine, 2026.

@article{veralopez2026respiratory,
  title   = {Improving respiratory disease detection through SSL-enhanced
             acoustic analysis and exercise-rest measurements},
  author  = {Vera-L{\'o}pez, {\'A}lvaro and Tilves-Santiago, Dar{\'i}o and
             Ram{\'i}rez-S{\'a}nchez, Jos{\'e} Manuel and Doc{\'i}o-Fern{\'a}ndez, Laura and
             Garc{\'i}a-Mateo, Carmen and Bustillo-Casado, Mar{\'i}a and
             Garc{\'i}a-Caballero, Alejandro Alberto},
  journal = {Frontiers in Medicine},
  year    = {2026},
  doi     = {10.3389/fmed.2026.1864436}
}

doi:10.3389/fmed.2026.1864436