COUGHVID Dataset: Cough Audio for ML Diagnostics

Updated 2 September 2025

COUGHVID dataset is a large-scale, crowdsourced collection of cough audio recordings with detailed demographic and clinical metadata, advancing respiratory research.
Its automated processing pipeline uses an XGB classifier and precise signal filtering to extract 68 acoustic features, ensuring accurate cough detection.
Expert clinical annotations and validation through GPS cross-referencing bolster the dataset's reliability for benchmarking ML models in COVID-19 and other respiratory conditions.

The COUGHVID dataset is a comprehensive, crowdsourced corpus of cough audio signals designed to advance the development and evaluation of machine learning algorithms for respiratory condition detection, especially COVID-19. Its large scale, global demographic diversity, clinically relevant annotations, and rigorous processing pipeline distinguish it as a foundational resource for audio-based health diagnostics.

1. Dataset Composition and Demographic Coverage

COUGHVID comprises over 20,000 cough recordings, collected globally via smartphone and web interfaces. Subjects self-report demographic attributes and health status, resulting in metadata with variables such as age (average 34.4 years, standard deviation 12.8 years), gender (65.5% male, 33.8% female, remainder "other"), and geolocation (when available). Health status categories include "healthy," "symptomatic," and "COVID-19 diagnosed," with approximately 1,010 recordings from self-declared COVID-19 positive individuals; 15.5% of labels are symptomatic, 7.5% COVID-19, and 77% healthy. Geographic distribution is confirmed by cross-referencing GPS data with public epidemiological statistics to validate COVID-19 prevalence in source regions (Orlandic et al., 2020).

2. Data Processing and Cough Detection Pipeline

A critical component of COUGHVID is its automated cough detection and audio preprocessing, ensuring the reliability of downstream ML applications. All recordings are filtered using an open-source eXtreme Gradient Boosting (XGB) classifier, pretrained to discriminate cough events from background noise and speech. Preprocessing steps:

Downsample audio to 12 kHz.
Lowpass filter with $f_{cutoff}$ = 6 kHz, formally: $x_{filtered}(t) = LPF\{x(t)\}$ .
Feature extraction: 68 acoustic features per recording, including those from Pramono et al., energy envelope peak detection, signal length, and power spectral density (PSD) across eight hand-selected frequency bands.

A classifier output in $[0,1]$ quantifies cough probability for each recording; only records with $cough\_detected \geq 0.8$ are retained for analytic use, while lower-probability samples are reserved for robustness tests (Orlandic et al., 2020).

3. Expert Annotations and Validation

Expert clinical labeling is a distinguishing feature of COUGHVID, addressing the challenge of unreliable self-reports. Three pulmonologists annotated over 2,000 recordings; each labeled 1,000 samples, with 15% overlap for inter-rater reliability assessment.

Annotations include:

Recording quality (good, ok, poor, or not a cough).
Cough type (wet, dry, indeterminate); audible findings such as dyspnea, wheezing, stridor, choking, and nasal congestion.
Clinical impressions: upper/lower respiratory infection, obstructive lung disease (e.g., COPD, asthma), COVID-19, healthy cough.
Severity: mild, severe, pseudocough.

Consistency is reported via Fleiss’ Kappa: nasal congestion ( $K_{Fleiss} \approx 0.49$ ), cough type ( $K_{Fleiss} \approx 0.23$ ), diagnostic impression (near-zero, reflecting clinical complexity). Stratified sampling ensured expert-labeled cases span all self-report categories. This provides rich, variable label quality, supporting research on label noise and aggregation (Orlandic et al., 2020).

4. Technical Methodologies for Machine Learning

COUGHVID was designed to facilitate the training and benchmarking of ML models for automated cough classification and COVID-19 diagnosis. Key methodological elements:

Temporal and spectral features (e.g., MFCCs, energy envelope, PSD bands) enable the use of both classical and deep learning models.
Feature extraction and filtering protocols are precisely defined (including LaTeX-formulated steps) to support reproducibility, e.g., energy envelope, spectral band power quantification.
Metadata integration (demographics, geolocation, health status, expert impressions) enables multivariate or multi-modal modeling pipelines.
Proven applicability for architectures such as convolutional neural networks (CNNs), XGBoost, and ensemble models. For example, deep models like CIdeR and transfer learning-derived feature representations have shown strong ROC-AUC metrics on similar datasets (Coppock et al., 2021, Fakhry et al., 2021).

5. Applications and Practical Utility

Beyond COVID-19 diagnosis, COUGHVID is applicable to a broad range of respiratory condition detection tasks:

Training and evaluation of ML models for cough event detection, wet/dry phenotype classification, and severity assessment.
Development of smartphone-based screening tools for scalable health surveillance.
Research in automated anomaly detection, symptom progression monitoring, and epidemiological modeling.
Data fusion with geolocation and temporal metadata enables outbreak surveillance and targeted population health interventions.

COUGHVID’s scale and annotation depth provide the basis for developing rapid, low-cost diagnostic solutions, especially in resource-limited or mass-testing settings (Orlandic et al., 2020).

6. Limitations, Labeling Challenges, and Mitigation Strategies

COUGHVID faces typical crowdsourcing constraints: inhomogeneous acoustic quality, device variability, self-report reliability, and inter-expert annotation disagreement. The dataset’s processing pipeline mitigates non-cough and noisy recordings via strict classifier thresholds; label refinement by experts addresses ambiguity.

Studies have proposed post-hoc semi-supervised learning strategies to reconcile expert disagreement and enhance label consistency, achieving up to threefold increases in class separability (measured by Jensen–Shannon divergence) and significant improvements in spectral discriminability (notably within 1–1.5 kHz frequency bands) (Orlandic et al., 2022). This robust labeling supports improved generalization and explainability for ML classifiers.

7. Impact, Research Adoption, and Subsequent Developments

COUGHVID has formed the empirical foundation for several subsequent research initiatives:

It has been incorporated into transfer learning pipelines and multi-dataset benchmarking frameworks, consistently improving cross-dataset generalizability (Islam et al., 2 Jan 2025).
The dataset catalyzed innovations in feature engineering for COVID-19 detection—including multi-branch deep learning systems, self-supervised transformer encoders, and robust ensemble classifiers achieving ROC-AUC scores within 0.93–0.99 on multiple datasets (Fakhry et al., 2021, Islam et al., 2 Jan 2025, Luong et al., 4 Aug 2025).
It is referenced as a benchmark for comparison in newer, clinically verified datasets (e.g., UK COVID-19 Vocal Audio Dataset) and in studies highlighting the necessity of global, diverse, and high-quality datasets for minimizing model bias (Budd et al., 2022, Haritaoglu et al., 2022).

A plausible implication is that COUGHVID’s scale and diversity will continue to drive advances in non-invasive, audio-based respiratory health screening, serve as a reference standard for future data collection protocols, and support research in robust label aggregation, domain adaptation, and multimodal diagnostics.

In summary, COUGHVID is a technically rigorous, large-scale, crowdsourced cough audio dataset with extensive demographic coverage and clinical annotation, designed to support the scientific paper and practical application of machine learning methods for respiratory health diagnostics, particularly COVID-19. Its validated preprocessing, expert annotation, and open-source character have made it a principal resource for both clinical research and algorithmic innovation in acoustic health monitoring (Orlandic et al., 2020).