A Preliminary Application of Echo State Networks to Emotion Recognition

This report investigates a preliminary application of Echo State Networks (ESNs) to the problem of automatic emotion recognition from speech. In the proposed approach, speech waveform signals are directly used as input time series for the ESN models, trained on a multi-classification task over a discrete set of emotions. Within the scopes of the Emotion Recognition Task of the Evalita 2014 competition, the performance of the proposed model is assessed by considering two emotional Italian speech corpora, namely the E-Carini corpus and the emotion corpus. Promising results show that the proposed system is able to achieve a very good performance in recognizing emotions from speech uttered by a speaker on which it has already been trained, whereas generalization of the predictions to speech uttered by unseen subjects is still challenging.