Emotion Recognition for Romanian Language Using MFSC Images with Deep-Learning Neural Networks

The objective of this paper is to study the emotion recognition for the Romanian language using MFSC images (Mel Frequency Spectral Coefficients) with DL-NN (Deep Learning Neural Networks). We obtained an accuracy of 85%, which is comparable to those reported by other researchers The emotions analyzed in this study are happiness, sadness, fury and neutral tone. The emotion recognition was tested in two cases: low and increased resolution MFSC images. We noticed that the MFSC images with higher resolution give better results. We have tested several architectures of DL-CNN (Convolutional Neural Network), but the best results were obtained when the network contains two autoencoder layers. We varied the number of neurons in these autoencoders in order to improve the emotion recognition accuracy for low / high resolution MFSC images.