Publication

Abstract : Effective automatic baby cry sound detection plays a significant role in many applications of smart baby condition monitoring. The paper presents deep learning based effective baby cry sound detection (BCSD) method under different kinds of background sounds in indoor environments. We investigated the functioning of the three BCSD approaches by making use of mel-frequency cepstral coefficients (MFCC) and machine learning classifiers like one-dimensional convolutional neural networks (1D-CNN), feed-forward neural networks (FFNNs), and multi-class support vectors machines (MC-SVM). We created the baby crying sounds for both training and testing of three models. In this study, we looked into the results of three methods under different frame lengths including 100 milliseconds, 250 milliseconds and 500 milliseconds. Evaluation results showed that the 1D-CNN with frame length of 500 MS provides promising results as compared to that of the frame lengths 100 milliseconds and 250 milliseconds. For frame length of 500 milliseconds, the 1D-CNN-based, FFNN-based and SVM-based BCSD method had F1-score of 98.86%, 98.46% and 97.97%, respectively for detecting cry sounds. The 1D-CNN based BCSD method had class-wise F1-score of above 98%. Results showed that three BCSD methods have promising results for the same test sound database including air-conditioner, fan, speech and music sounds.