Analysis of Respiratory Health Indicators in Speech-Breathing-Patterns

Speech production and respiration co-occur, and hence, speech signals encode information regarding an individual's respiratory health. Disorders within the respiratory tract often diminish lung capacity, thereby affecting speech production mechanisms. In this study, we leverage breath patterns extracted from speech to assess inhalation and exhalation capacities during speaking. Hence, the area under the breath pattern curve (AUBPC) derived from speech signals emerges as an indicator of lung capacity. We propose utilising AUBPCs as one of the parameters indicating respiratory disorders in speech signals from the Coswara dataset. To extract breath patterns from speech, we employ SBreathNet, a pre-existing model trained on speech signals to generate breath patterns as output. The breathing patterns (BPs), serving as a representation technique, demonstrates the ability to discern nine distinct respiratory conditions with an absolute AUC increase of 6% as compared to mel-frequency cepstral coefficients (MFCCs) on the same dataset.