Cross-cultural Emotion Classification: the Effect of Emotional Intensity and Acoustic Features

Anonymous

Abstract: Cross-cultural emotion recognition is attracting increasing research attention; robustness to such differences in emotional expression is important for speech modality emotion recognition. In this work we quantify the accuracy loss when classifying cross-culturally for multiple emotional intensities, and investigate the effect of feature sets, including feature importance. We find that different emotional intensities yield a similar decrease in cross-culture accuracy relative to within-culture, and different acoustic feature sets also yield similar relative cross-culture accuracy. The top 10 important eGeMAPS features for within-cultural and cross-cultural classification share only one common feature, which partially explains differences in accuracy.