SpiroPhonia: Non-Invasive Respiratory Health Assessment from Spontaneous Speech

1. Introduction

• Clinic-bound spirometry requires clinical settings, trained personnel, and patient cooperation, leaving early disease underdiagnosed and monitoring episodic.These constraints motivate scalable, passive, at-home assessment modalities.Source: §1 · p. 1
• Spontaneous speech needs no prescribed task or controlled setting, but variability in content, prosody, speakers, and recording conditions may obscure respiratory biomarkers.The paper asks whether such speech remains discriminative under uncontrolled conditions.Source: §1 · pp. 1–2
• 78% accuracy, 80% F1-score, and 87% AUC were achieved by SpiroPhonia on a 201-speaker dataset under subject-independent evaluation.The dataset included 102 speakers with COPD or related respiratory conditions and 99 healthy controls.Source: §1 · p. 2

2.1. SpiroPhonia Dataset Curation and Collection

• 201 speakers were assembled from publicly available interviews and narratives, with respiratory-status labels verified by a physician and 99 healthy controls.The dataset included 102 speakers with respiratory conditions; recordings were manually checked for audio quality and conversational content.Source: §2.1 · p. 2

2.2. Preprocessing

• Speech was cleaned, divided into 10–30-second windows, amplitude-normalized, band-pass filtered from 100 to 5000 Hz, and lightly denoised when necessary.The window range balances pause-feature stability against topic- and style-related prosodic variability.Source: §2.2 · p. 2, Figure 2 · p. 2

2.3. Speech Feature Extraction and Analysis

• Features captured vocal stability, spectral structure, and breath–speech timing through acoustic perturbation measures, 13 MFCCs with five summary statistics each, and pause measures.Pause features included silence fraction, pauses per minute, and total pause duration.Source: §2.3 · pp. 2–3

2.4. Statistical Analysis and Summary of Feature Behavior

• p < 0.001: jitter parameters showed the strongest between-group differences; MFCC effects reached d ≈0.6, and all three pause features differed significantly.Pause ratio and pauses per minute had the highest rank-biserial correlations; Figure 3 presents representative distribution shifts.Source: §2.4 · p. 3, Figure 3 · p. 3

3.1. Model Design and Feature Optimization

• The two-stage optimization reduced 94 attributes to 67 non-redundant features, then selected classifier-specific subsets with RFECV.Feature processing and model fitting were restricted to the training data; the 41-sample held-out set was reserved for final evaluation.Source: §3.1 · p. 3

3.2. Classification Performance

• 78.05% accuracy was achieved by both Linear SVM and Gradient Boosting; Random Forest had the highest AUC at 87.14%.SVM specificity was 85.00% using five features, while Gradient Boosting sensitivity was 85.71% (95% CI: 68.74–100.00%).Source: §3.2 · p. 4

3.3. Discriminative Speech Biomarkers

• Linear models, especially SVM, performed strongly with five features, while Logistic Regression and the shallow Neural Network converged around 14.Random Forest and Gradient Boosting retained roughly 59 predictors; SVM accuracy stabilized beyond approximately ten selected features.Source: §3.3 · p. 4, Figure 4 · p. 4
• Local jitter, local shimmer, and pause ratio were selected by every model, alongside frequently selected mid-order MFCCs.The reported MFCC groups include MFCC2–MFCC5 and MFCC10–MFCC12.Source: §3.3 · p. 4

3.4. Clinical Utility Assessment

• Gradient Boosting is suited to screening, Random Forest to high-confidence detection, and five-feature Linear SVM to confirmatory or resource-constrained deployment.These roles reflect their reported sensitivity, AUC, and specificity profiles, respectively.Source: §3.4 · p. 4

3.5. Comparative Evaluation

• 78.05% accuracy was achieved by compact feature sets on spontaneous speech, comparable to prior controlled-setting approaches.The cited comparison reports prior accuracies of 67%–84.6%.Source: Table 5 · p. 4, §3.5 · p. 4

• Spontaneous conversational speech contains acoustic and prosodic signatures that support subject-independent discrimination between COPD and healthy speakers under real-world recording variability.Source: §4 · p. 4
• A compact, interpretable predictor set reflects phonatory instability, altered spectral structure, and disrupted breath-speech coordination.The conclusion describes competitive performance with minimal features as supporting the feasibility of respiratory assessment from everyday speech.Source: §4 · p. 4
• The modest dataset limits evaluation of more data-intensive models, and broader language, device, and acoustic-environment validation remains needed.Source: §4 · p. 4

• Generative AI tools were used solely for language editing and polishing; the authors state that they conceived and performed the scientific work.Source: §6 · p. 5