Deep Learning of Attitude in Children’s Emotional Speech

Attitude comprises a paralinguistic information associated to affective states without the explicit connection to either positive nor negative valence. As such, its automatic recognition plays an integral role in all speech-based systems. This work focuses on classifying attitude in children’s emotional speech into four classes, i.e. confident, uncertain, apathetic, and enthusiastic. We present a flexible classification scheme based on a directed acyclic graph able to easily incorporate heterogeneous feature sets and classifiers according to the necessities of each task. More precisely, this work employs features coming from both frequency and wavelet domains combined with convolutional and recurrent neural networks. The obtained results confirm the efficacy of such a classification graph outperforming traditional neural network schemes.