Regression estimation model for emotion and intensity of speech using perception rating

Abstract

An emotional intensity regression estimation model was created using calculated perceived intensity values and deep learning. In our previous study, we considered emotional intensity using 10 categories and estimated emotional intensity by categorization, but the flexibility of the method was insufficient. In order to solve this problem, an emotional intensity estimation model which takes into account differences in the perceptual intensity value of each category of emotional intensity was used in this study. For this purpose, two types of perceived intensity values were calculated for a Japanese speech corpus of sounds uttered in an emotion perception rating experiment. In the results, the average correlation coefficient between the estimated intensity value and the set intensity value of the sounds was 0.73 for the emotional intensity estimation model when perceived intensity values were used. These results suggest the possibility of successfully estimating emotional intensity using regression.