An End-to-end System Based on TDNN for Lung Sound Classification

In this paper, an efficient lung sound classification system based on the end-to-end neural network is proposed to classify the normal and abnormal lung sounds. We realize the time-delay neural network (TDNN) to process the acoustic characteristics of lung sounds, and make the classification on the output layer of TDNN with Softmax function to obtain the relative prediction probability. In addition, we explore three different acoustic features, such as Mel-frequency ceptral coefficients (MFCC), inverted Mel frequency cepstral coefficients (IMFCC), log Mel-filterbank energies (FBank), to study the potential impact on the system performance. Finally, we evaluate the robustness of the proposed end-to-end system under noisy conditions. The experimental results showed that the proposed end-to-end lung sound classification system achieved the outstanding performance in noisy environment.