COUGHVID: Cough Audio Dataset
- COUGHVID is a large, publicly available crowd-sourced dataset that compiles over 20,000 cough recordings with self-reported metadata and expert annotations.
- It supports a range of tasks including cough detection, COVID-19 classification, wet-versus-dry cough analysis, and demographic inference using both automatic and manual labeling.
- The dataset underscores challenges such as label noise, inter-rater variability, and dataset shift, serving as a benchmark for robust respiratory-audio ML research.
COUGHVID is a large-scale, publicly available, crowdsourced cough-audio dataset introduced to support machine learning research on cough analysis, respiratory disease characterization, and especially COVID-19 screening from cough sounds. The original release contains more than 20,000 cough recordings collected worldwide through a web application hosted on an EPFL server, while later work used an expanded version with about 34,500 recordings collected from April 2020 to October 2021, of which 20,644 have user status labels. Across these versions, COUGHVID combines raw audio, self-reported metadata, cough-detection scores, and limited expert annotation, and it has become a recurring benchmark for cough detection, COVID-19 classification, wet-versus-dry cough characterization, symptom and abnormality detection, relabeling studies, and cross-dataset generalization analysis (Orlandic et al., 2020).
1. Origin, scope, and dataset design
COUGHVID was created in response to a specific bottleneck in respiratory-audio ML: the lack of a sufficiently large, validated, publicly available cough database for training robust models. The dataset was designed to support cough/non-cough detection, COVID-19 screening, detection of other respiratory conditions, cough quality assessment, and broader studies of cough acoustics and user metadata. In the original release, recordings were collected between April 1, 2020 and September 10, 2020 through a web interface built around a “one recording, one click” principle; users could record microphone audio for up to 10 seconds and then optionally provide age, gender, respiratory condition, fever/muscle pain, self-reported COVID-related status, and geolocation (Orlandic et al., 2020).
The resulting corpus is intrinsically heterogeneous. Recordings were collected from users around the world, with broad variation in age, gender, health status, respiratory conditions, and acquisition conditions. The original paper reports more than 20,000 cough recordings, including 1,010 self-reported as COVID-19. Among recordings with metadata and cough probability above 0.8, 65.5% were male, 33.8% female, 81.9% reported no pre-existing respiratory condition, 77% were healthy, 15.5% symptomatic, and 7.5% COVID-19 positive; the mean age was 34.4 years with standard deviation 12.8 years. This mixture of scale, optional metadata, and crowd acquisition is central to COUGHVID’s later role as both a useful resource and a difficult benchmark.
The dataset structure reflects that dual role. Each recording is stored as an audio file in .webm or .ogg encoded with Opus at 48 kHz together with a corresponding .json metadata file, and a compiled metadata_compiled.csv with 40 columns and one row per record is provided. The metadata span context information such as timestamp and cough detection probability, self-reported user information, and expert annotations. The original authors explicitly note that, aside from datetime and cough_detected, every variable may be considered a prediction target, which makes the corpus relevant not only for COVID-19 detection but also for dry-versus-wet cough classification, symptom inference, diagnosis prediction, severity estimation, and demographic inference (Orlandic et al., 2020).
2. Acquisition pipeline, filtering, and annotation structure
Because crowdsourced respiratory audio contains substantial non-cough material, COUGHVID incorporates an explicit filtering stage. The original dataset paper describes an automatic cough detector trained on 121 cough sounds and 94 non-cough sounds, with the non-cough set including speech, laughter, silence, and background noise. Audio for detector training was lowpass filtered at kHz and downsampled to 12 kHz, and the detector used 68 audio features: 40 features from Pramono et al., 19 energy envelope peak detection features from Chatrzarrin et al., 1 signal-length feature, and 8 PSD features from hand-selected frequency bands. An XGBoost classifier was tuned with Tree-structured Parzen Estimators using a precision objective and 10-fold, 20%-test shuffle-split cross-validation. The resulting cough_detected score functions both as a filter and as a quality indicator; recordings with cough_detected < 0.8 are mostly not valid coughs, and only 10.4% of them actually contain cough sounds (Orlandic et al., 2020).
The reported performance of this detector is substantial but not perfect: 89.22% cross-validation precision, 89.34% testing balanced accuracy, 96.5% ROC, 90.56% F1-score, and 88.89% testing precision. In the original release, 10,743 recordings with cough probability above 0.8 were considered cough sounds, and 6,485 of those had GPS information. The authors further performed a plausibility check linking self-reported COVID/symptomatic recordings to countries with active transmission; 94.4% of COVID-labeled recordings and 91.3% of symptomatic recordings came from countries with more than 20 new confirmed cases per 1 million people in the prior 14 days. This was presented as a plausibility check rather than clinical confirmation.
Clinical annotation was added through expert pulmonologist review. In the original release, three expert pulmonologists each labeled 1,000 recordings, with 150 recordings labeled by all three experts to assess agreement, yielding more than 2,000 expert-labeled recordings overall. Recordings selected for expert labeling were prefiltered by the cough detector with probability at least 0.8 and then sampled in a stratified random way based on self-reported status: 25% COVID, 35% symptomatic, 25% healthy, and 15% with no status reported. Each expert answered 10 annotation items spanning recording quality, cough type, audible dyspnea, wheezing, stridor, choking, nasal congestion, “nothing specific,” diagnostic impression, and severity impression. The labeling workflow used Google Sheets with in-browser playback, and each expert spent about 10 hours labeling 1,000 recordings (Orlandic et al., 2020).
These labels made COUGHVID unusually versatile. The expert-labeled COVID subset in the original pooled analysis comprised 632 total COVID-19-labeled cough records after combining all expert labels and resolving overlaps by prioritizing Expert 1 in overlap cases. The pooled COVID-labeled subset was described as relatively homogeneous in some respects: 87.3% labeled dry, 86.2% labeled mild, 93.0% without audible dyspnea, 90.5% without wheezing, 98.7% without stridor, 99.1% without choking, and 99.2% without nasal congestion. A private test set, drawn from recordings labeled by at least one expert, was also retained to support external benchmarking (Orlandic et al., 2020).
3. Label noise, inter-rater disagreement, and relabeling efforts
A defining feature of COUGHVID is that scale was achieved through crowdsourcing rather than controlled clinical acquisition, and later work repeatedly emphasized the consequences for label quality. The original overlap analysis among the three pulmonologists found poor agreement for several key labels. Fleiss’ kappa was for quality, $0.23$ for cough type, $0.04$ for dyspnea, $0.04$ for wheezing, for stridor, for choking, $0.49$ for congestion, $0.10$ for “nothing,” and effectively zero for diagnosis, with for the diagnosis label. Only 22 of 86 coughs that at least one rater labeled as COVID-19 had majority consensus (Orlandic et al., 2020).
Later semi-supervised relabeling work made this limitation explicit and treated COUGHVID itself as a case study in annotation inconsistency. Using an expanded version with about 34,500 recordings and 20,644 user status labels, that study described three major issues: user mislabeling, expert disagreement, and label sparsity. Four physicians had each labeled 1,000 recordings for audible respiratory disorders, with a shared subset of 150 recordings used to assess inter-rater agreement. The reported overall Fleiss’ kappa was 0, and Expert 3 labeled only one recording as COVID-19, making that expert unsuitable for a standalone model. The authors therefore proposed a semi-supervised learning pipeline based on expert-specific models, pseudo-label propagation, and agreement filtering to recover a higher-confidence subset (Orlandic et al., 2022).
That relabeling framework used majority agreement as the final selection rule. It labeled 1,018 previously unlabeled recordings, relabeled 581 “symptomatic” user-labeled recordings into COVID-19 or healthy, and produced a subset whose average Jensen–Shannon divergence rose from 0.00877 for the user-labeled data to 0.0284 for the SSL majority-agreement data, more than 3× higher separability. The selected SSL training set was only 23% smaller than the user-labeled one. The same study reported significantly larger bandpower differences between healthy and COVID-19 coughs in the relabeled subset, specifically in the 400–550 Hz range with 1 and in the 1000–1500 Hz range with 2, and achieved a private-test AUC of 0.797 for the final SSL model versus 0.562 for the user-label model and 0.681, 0.743, and 0.593 for the three single-expert models. The new labels were added to the public repository in a status_SSL metadata column (Orlandic et al., 2022).
These findings established a recurring interpretation of COUGHVID: it is not merely a large cough repository but also a dataset in which annotation uncertainty is itself a research problem. A plausible implication is that performance gains on COUGHVID may reflect not only model quality but also the degree to which a study controls, repairs, or exploits label inconsistency.
4. Roles in downstream tasks and benchmark protocols
COUGHVID has been used in several distinct experimental roles. In the original framing, it supported tasks such as cough detection, diagnosis prediction, and severity estimation. Subsequent work specialized these uses. One important direction treats COUGHVID as a COVID-19 classification benchmark. Another uses it for clinically meaningful cough characterization. A third uses it as an external data source to compensate for class imbalance in other datasets.
A notable example of the second direction is CoughViT, which uses COUGHVID for wet-or-dry cough classification. In that study, COUGHVID serves as a benchmark for a clinically relevant cough characterization task rather than only for COVID-19 detection. The benchmark uses Expert 4’s annotations because of low inter-rater agreement, and the model is pretrained self-supervised on COVID-19 Sounds before evaluation on COUGHVID, COVID-19 detection, and cough detection. On the COUGHVID wet-or-dry task, CoughViT achieved 74.95 AUROC in the main fine-tuning setting and 0.71 AUROC on the COUGHVID blind test set, compared with 0.59 for a logistic regression baseline and 0.56 for AST-Audioset (Luong et al., 4 Aug 2025).
COUGHVID has also been used as an auxiliary positive-class source in challenge settings. A ResNet-50 DiCOVA system described COUGHVID as an open-source database of COVID-19 coughs collected by the EPFL CoughVid team and reported that the full pool contained 20,072 audio samples, of which 1,010 were from patients diagnosed as COVID-19 positive. That study used the dataset’s cough probability field and COVID-confirmed status to filter the corpus and extracted 640 additional COVID cough samples for training, explicitly to increase the number of COVID-positive samples and enhance variability in the training data (Banerjee et al., 2021).
Benchmarking studies have also used COUGHVID as a stress test for reliability and dataset shift. Sound-Dr treated COUGHVID as one of two external respiratory-sound datasets for comparison and highlighted several practical issues: some COUGHVID samples are less than 1 second long, these short recordings may be unqualified for training, they can also cause errors in reading input data, and the reported COUGHVID sampling rate in that benchmarking setup was 22050 Hz versus 48000 Hz for Sound-Dr. In the same work, COUGHVID showed weak supervised COVID-19-detection results under a common feature-and-SVM pipeline, with AUC values around 53–55, and unsupervised abnormality-detection results that were also weak relative to Sound-Dr (Hoang et al., 2022).
Taken together, these studies show that COUGHVID is not tied to a single task definition. It functions as a general cough-analysis corpus, a COVID-19 screening benchmark, a wet-versus-dry cough benchmark, an auxiliary source for data augmentation, and a dataset-shift probe.
5. Reported performance and protocol dependence
A striking feature of the COUGHVID literature is the wide dispersion of reported results. Under some controlled in-dataset protocols, performance is high. In other studies, especially those imposing stricter balancing or cross-dataset transfer, performance is only modest. This variation is not incidental; it is directly tied to differences in label definitions, filtering, class balance, demographic controls, feature pipelines, and evaluation design.
One line of work reports strong COUGHVID results under curated feature-engineering and model-selection pipelines. A 2025 cross-datasets study treated COUGHVID as a balanced binary classification task with 1,360 total cough samples, split into 680 COVID-19 positive and 680 COVID-19 negative recordings. Using a 193-dimensional handcrafted feature vector, RFECV with Extra-Trees, Bayesian optimization, SMOTE, threshold moving, and a Deep Neural Decision Forest, it reported 0.93 accuracy, 0.93 AUC, 0.93 precision, 0.94 recall, 0.93 F1-score, and 0.93 specificity on COUGHVID. The same paper noted that earlier strategy variants were much weaker, with COUGHVID AUC progressing from 0.63–0.65 in the default setting to 0.92–0.93 only after the full tuning pipeline (Islam et al., 2 Jan 2025).
Another study focused on acoustic feature comparisons and used a filtered subset of 428 COUGHVID recordings, segmented into 1,719 cough segments after preprocessing. In that setup, the best-performing configuration was MLP plus MFCC, with 84.30% accuracy and AUC 0.843. The same paper reported weaker COUGHVID performance for Chroma and Spectral Contrast features and attributed the gap relative to Virufy partly to background noise, crowdsourced recording variability, and microphone limitations, especially for capturing low frequencies (Shati et al., 2023).
By contrast, a technical report on demographic stratification described COUGHVID as originally containing 34,434 recordings, then filtered it to samples with valid age and gender and excluded the symptomatic class. After stratified undersampling by age × gender, the final COUGHVID set contained 2,160 samples, with 1,080 COVID-19 and 1,080 healthy recordings. Under this protocol, intra-dataset results were modest: Audio-MAE reached 0.60 AUC, while CNN6 achieved 0.63 AUC and 0.67 F1. The same report found that fine-tuning Audio-MAE on the unbalanced version increased AUC from 0.60 to 0.63 and interpreted that increase as evidence of demographic leakage rather than true acoustic learning. Cross-dataset evaluation involving COUGHVID was poor in both directions, with AUC values ranging from 0.43 to 0.68 and especially severe failure for Audio-MAE in some transfer settings (Brito et al., 18 Nov 2025).
This protocol dependence has become one of the central technical lessons of COUGHVID. Strong in-dataset scores can be obtained, but they coexist with repeated evidence that stricter control of confounders and transfer conditions can substantially lower performance. A plausible implication is that COUGHVID is best understood as a benchmark whose difficulty depends heavily on how one defines the label space and controls nuisance structure.
6. Limitations, scientific significance, and continuing role
The main limitations of COUGHVID are consistent across the literature. The dataset is crowdsourced, so recording quality, environmental noise, and device characteristics vary substantially. Self-reported labels are inherently noisy. Metadata are optional and therefore sparse in places. Expert review covers only a limited subset, and inter-rater agreement on diagnosis is poor. Some downstream studies also emphasize short clips, sampling-rate heterogeneity in derived benchmark versions, class imbalance, and missing demographic metadata as practical obstacles (Orlandic et al., 2020).
At the same time, those limitations are inseparable from COUGHVID’s value. It is a large, open, metadata-rich cough corpus collected “in the wild,” and later work repeatedly used it to study precisely the issues that hinder real deployment: label inconsistency, dataset shift, demographic confounding, and the gap between in-domain benchmarking and cross-domain robustness. In that sense, COUGHVID has functioned both as a resource and as a diagnostic instrument for the field. Studies on SSL relabeling, self-supervised cough representation learning, robustness benchmarking, and demographic balancing all treat the dataset as a testbed for methods intended to cope with noisy, partially labeled respiratory audio rather than as a clean clinical reference (Orlandic et al., 2022).
Its continuing significance is therefore methodological as much as empirical. COUGHVID helped establish that large-scale cough collection is feasible, that expert-labeled subsets can support clinically meaningful tasks such as dry-versus-wet cough characterization, and that downstream performance is highly sensitive to annotation quality and evaluation protocol. It also made visible a broader distinction in respiratory-audio research between large crowdsourced corpora and clinically validated datasets with RT-PCR-confirmed labels. This suggests that COUGHVID’s long-term importance lies not only in any single benchmark number, but in its role as a canonical dataset for studying how cough-audio models behave under realistic noise, ambiguity, and distribution shift.