A Cough Detection Method Based on the Conformer-BiLSTM Model

Cough is a common symptom of respiratory disease, and its detection is a basic step in cough sound analysis. Manual cough sound segmentation is tedious, subjective, and inefficient. Cough sounds from real-world scenarios can be collected in various environments using different devices, making cough event detection more difficult. This paper proposes a hybrid model of conformer and bidirectional long short-term memory (BiLSTM) networks to address this issue. It obtains contextual information with the BiLSTM model, extracts global features using the conformer model with multi-head attention residual connections, establishes comprehensive dependencies on the entire time series, and localizes the start and end times of each cough event. This method was trained and tested using 500 audio files containing 1928 cough sounds manually annotated from the Vocalsound audio public dataset. The results showed that a 97.22% sensitivity and 97.08% specificity were achieved. This method also achieved a sensitivity of 98.45%, specificity of 92.7%, and average time overlap rate of 81.12% in the clinical dataset.