Infant cry classification by MFCC feature extraction with MLP and CNN structures

In this study, Dunstan's infant cry data set is pre-processed with the feature vector approach, including MFCC (19 features) and energy (one feature). By using extracted features and Support Vector Machine (SVM), Multilayer Perceptron (MLP), and Convolutional Neural Network (CNN) classifiers, five classes of infant cry ("Neh" = hungry; "Eh" = need to burp; "Owh" = tired; "Eairh" = stomach cramp; "Heh" = physical discomfort) are distinguished. The proposed MLP and CNN structures are analyzed according to the loss and the accuracy based on the epoch; moreover, to evaluate the performance of classifiers AUC-ROC, Confusion matrix, accuracy, f1_score, recall, and precision have been used. All three classifiers are analyzed, and their results show that the CNN-designed model has the best performance. Results show that the performance will improve by increasing the complexity of the model. With this approach, classifiers are run 10 times, and the average accuracy for SVM for SMOTE and non-SMOTE data are obtained with tolerance 0.823 +/- 0.02, 0.861 +/- 0.02, respectively. These accuracies for MLP are 0.876 +/- 0.01, 0.892 +/- 0.01, and finally, for CNN, are 0.921 +/- 0.005, 0.911 +/- 0.005. At the best condition, an accuracy of 92.1 % is obtained for five classes of infant cries by the proposed CNN structure.