Emotion Recognition for Romanian Language using Deep Learning with DL-CNN convolutional layers

Abstract

In this paper, we used the SRoL (Voiced Sounds of the Romanian Language) emotional corpus for Romanian language. We used a Deep Learning Neural Network (DL-CNN) to automatically recognition four emotions: joy, sadness, fury and neutral tone. A 3-layer deep learning neural network is used (two autoencoders are associated to the first two hidden layers and the third layer is a softmax layer). The best results of the emotion recognition obtained is 84,48% with fine tunning and 74.95% without fine tunning. It was noted that the percentages of emotion recognition provided by the MFSC - mel-frequency spectral coefficients are better than those obtained by the MFCC - mel frequency cepstral coefficients.