Emotional Intensity Estimation of a Japanese Speech Corpus Using Acoustic Features

Recently, there have been many studies of emotion estimation using non-linguistic speech data. In addition, there have been a few studies on emotional intensity. However, failure to read this emotional intensity correctly can lead to errors in the responses humans and machines should make when communicating with each other. In this paper, we developed three models for emotional intensity estimation using deep learning, and examined the accuracy of emotional intensity estimation of a Japanese speech corpus, which resulted in 52.4% accuracy of emotional intensity estimation. We also investigated the correlations between acoustic features and analyzed the properties of acoustic features in order to improve estimation accuracy, and found that the differentiation of gammatone cepstral coefficients varied significantly between intensities.