Breathing features are valuable tools for detecting and diagnosing respiratory diseases from speech. The current study analyzed speech samples of interviews from 23 COVID-19 patients via social media platforms. Breathing features were extracted for each breath group (BG) locally and for each discourse globally. Perceptual tasks based on written texts and audio samples were conducted at both breath-group and discourse levels. PCA reduced breathing features to 1) at breath-group level: BG length, proportion of pauses, and speech rate; 2) at discourse level: discourse length, average BG duration, average inter-breath-group pauses (IBP) duration, and proportion of IBP to BG each within a discourse. Linear models showed that at breath-group level, speaker’s BG length can be positively predicted by the perceived text valence and negatively by the perceived text fluency, while proportion of pauses within BG negatively by fluency. At discourse level, the average BG duration has a negative predictive effect on the perceived probability of illness; additionally, the longer BG duration predicts higher illness severity, while the higher proportion of IBP to BG predicts lower illness severity. Our study presents a data-driven approach of breathing features associated with respiratory diseases and demonstrates the way how these features interact with speech perception.