Toward improving estimation accuracy of emotion dimensions in bilingual scenario based on three-layered model

This paper proposes a newly revised three-layered model to improve emotion dimensions (valence, activation) estimation for bilingual scenario, using knowledge of commonalities and differences of human perception among multiple languages. Most of previous systems on speech emotion recognition only worked in each mono-language. However, to construct a generalized emotion recognition system which be able to detect emotions for multiple languages, acoustic features selection and feature normalization among languages remained a topic. In this study, correlated features with emotion dimensions are selected to construct proposed model. To imitate emotion perception across languages, a novel normalization method is addressed by extracting direction and distance from neutral to other emotion in emotion dimensional space. Results show that the proposed system yields mean absolute error reduction rate of 46% and 34% for Japanese and German language respectively over previous system. The proposed system attains estimation performance more comparable to human evaluation on bilingual case.