Prediction of emotional states in parent-adolescent conversations using non-linear autoregressive neural networks

This study investigates an application of nonlinear autoregressive (NAR) models to the prediction of the most likely time series of emotional state transitions of speakers engaged in dyadic conversations. While, previous methods analyzed each speaker in separation, the new approach proposes to couple both speakers into a nonlinear recursive predictive neural network system (NARX-NN). The NARX-NN system was tested and compared with its uncoupled version (NAR-NN). The tests were conducted using speech recordings from 63 parent-child dyads including 29 depressed and 34 non-depressed adolescent children, 14-18 years of age. The conversations were conducted on three different topics. The NARX-NN outperformed the NAR-NN method in all experimental scenarios and across all topics of conversation. Predictions of emotional states for depressed children led to higher accuracy than the predictions for non-depressed children. Modeling with class and/or speaker dependency improved the results compared to the class and/or speaker independent models.