Constant-Q Based Harmonic and Pitch Features for Normal vs. Pathological Infant Cry Classification

Abstract

Classification of normal vs. pathological infant cry is a socially relevant and challenging problem. Many feature sets, such as Mel Frequency Cepstral Coefficients (MFCC), Linear Frequency Cepstral Coefficients (LFCC), and Constant Q Cepstral Coefficients (CQCC) have been used for this task. However, an effective representation of the spectral and pitch components of a spectrum together is not achieved leaving scope for improvement. Also, the infant cry can be considered a melodic sound implying that the fundamental frequency and timbre-based features also carry vital information. This work proposes Constant Q Harmonic Coefficients (CQHC), and Constant Q Pitch Coefficients (CQPC) extracted by the decomposition of the Constant Q Transform (CQT) spectrum for the infant cry classification. This work uses Convolutional Neural Network (CNN) as the classifier along with traditional classifiers, such as Gaussian Mixture Models (GMM) and Support Vector Machines (SVM). The results using the CNN classifier are compared by considering the MFCC, LFCC, and CQCC feature sets as the baseline features. The feature-level fusion of MFCC with log-CQHC and MFCC with log-CQPC achieved a 5-fold accuracy of 98.73% and 98.96% respectively, surpassing the baseline MFCC. Furthermore, the fusion of MFCC with log-CQHC and log-CQPC feature sets resulted in improved classification accuracy of 3%, 4.7%, and 5.85% when compared with the baseline MFCC, LFCC, and CQCC feature sets, respectively. Further, our intensive experiments using three classifiers structures, namely, GMM, SVM, and CNN indicate superior results using the proposed feature extraction techniques.