Classification of infant cries using source, system and supra-segmental features

In this paper, source, system and supra-segmental features are explored for recognition of infant cry. Different types of infant cries considered in this work are hunger, pain and wet-diaper. In this work, mel-frequency cepstral coefficients (MFCC), residual MFCC (RMFCC), implicit LP residual features, features from modulation spectrum and time domain envelope features are used for representing the infant cry specific information from the acoustic signal. Gaussian Mixture Models (GMM) are used for classifying the above mentioned cries from the features proposed in this work. GMM models are developed separately by using the proposed features. Infant cry database collected under telemedicine project (eNPCS) at IIT-KGP has been used for carrying out this study. The recognition performance of the developed GMM models is observed to be varying significantly based on the features. Results have indicated that, the proposed features have complementary evidences in view of discriminating the infant cries. For enhancing the recognition performance, GMM models developed using various features are combined using score level fusion. The recognition performance using combination of evidences is found to be superior over individual systems.