A Speech Dataset for the Recognition and Analysis of COVID-19 Coughs

The coronavirus and its variants have spread globally, and since the virus primarily transmits through the air, efficient and convenient self-testing methods have become crucial for controlling the pandemic. Current COVID-19 detection technologies generally rely on data-intensive deep learning frameworks; however, the scarcity of COVID-19 cough voice samples has been a major barrier to advancement in this field. To address this challenge, this study innovatively constructed a specialized dataset comprising 3,395 audio recordings, covering three core categories: COVID-19 coughs, non-COVID-19 coughs, and healthy breathing sounds, ensuring data comprehensiveness and specificity. To further explore the deep features of sound signals, we employed four cutting-edge signal processing techniques—Fast Fourier Transform (FFT), Short-Time Fourier Transform (STFT), Mel-Frequency Cepstral Coefficients (MFCC), and Continuous Wavelet Transform (CWT)—to extract the spectral and time-frequency characteristics of the sounds, generating four sets of feature-rich images. Based on this, we selected 12 state-of-the-art classification models and conducted a comprehensive experimental process to establish a performance benchmark. The experimental results demonstrated that most models exhibited excellent recognition accuracy on our dataset, confirming both the high quality of the dataset and the effectiveness of the models.