Cough Sound Analysis using Vocal Tract Models

Coughing is one of the most common symptoms of respiratory disorders, and variations in the cough sound relate to factors such as the type and quantity of secretions, physiological differences in the airway, and the force of the expulsion of the air. Automated analysis of spontaneous cough sounds in a smart home can provide a non-contact, non-invasive method to identify changes in health status, and passively monitor the progress of conditions such as chronic lung disease, to support independence and aging in place. In this work we propose analyzing and characterizing coughs using vocal tract models originally developed for speech coding. These models can relate cough sounds to physiological features, helping provide the interpretable and explainable predictions that are necessary for trust and confidence in healthcare applications. We show that linear prediction can effectively capture the time and frequency dynamics of different cough phases. We also demonstrate the interpretability of the model parameters by developing features to distinguish wet and dry cough sounds. The features achieve perfect linear separation of coughs in a small physician-labelled dataset and provide insight into the sound characteristics that contribute to those descriptors. The compact representation of cough sounds provided by the parametric model approach motivates further investigation for embedded cough analysis applications.