Cough audio signal classification has been successfully used to diagnose a variety of respiratory conditions, and there has been significant interest in leveraging Machine Learning (ML) to provide widespread COVID-19 screening. The COUGHVID dataset, collected by the EPFL between April and December 2020, provides over 25,000 crowdsourced cough recordings representing a wide range of participant ages, genders, geographic locations, and COVID-19 statuses.
In addition to the raw audio recordings, we have undergone an additional layer of validation whereby four expert physicians annotated a fraction of the dataset to determine which crowdsourced samples realistically originate from COVID-19 patients. In addition to COVID 19 diagnoses, the expert labels and metadata provide a wealth of insights beyond those of existing public cough datasets, which either do not provide labels or contain a small number of samples. The COUGHVID dataset publicly contributes over 2,800 expert-labeled coughs, all of which provide a diagnosis, severity level, and whether or not audible health anomalies are present, such as dyspnea, wheezing, and nasal congestion. Using these expert labels along with participant metadata, the dataset can be used to train models that detect a variety of participants’ information based on their cough sounds, also beyond COVID-19.
Finally, along with the dataset an open-source cough detection ML model is provided to help filtering non-cough recordings from the database. This automated cough detection tool assists developers in creating robust applications that automatically remove non-cough sounds from their databases. Furthermore, this open-sourced cough detection model, preprocessing methods, cough segmentation algorithm, and SNR estimation code enable the research community to seamlessly integrate their own datasets with COUGHVID while keeping the data processing pipeline consistent.