Child emotion detection using prosodic features with neural network based model

This paper focuses on the detection of child emotions in speech through classification of regional Oriya speakers. In this paper, an approach has been taken to detect the child emotion using prosodic features. As prosody of the children is more informative, different features i.e. energy, autocorrelation and pitch are considered in this work. In this study, we detect three emotional classes: Anger, Fear, and Happy. The popular neural network based classifier as multiple linear Perceptron (MLP) and radial basis function network (RBFN) is chosen for this task. The result of the proposed method for classification is found as79.6%, 84.6% and 92.9% for proposed sets of features correspondingly. An emotional speech database from children of Indian regional Oriya language has been developed, tested and compared with similar databases used in this area. Comparison proves that Mel frequency cepstral coefficients (MFCC) features are found one of the best features for detection from speech. The database comprises of Indian regional Oriya language emotional children speech utterances exhibited in natural environment and is first of its kind database in the chosen language.