A New Look at Breathing for Affective Studies

Abstract

In affective computing, breathing has seen lighter use than the heart and EDA channels. Several reasons have contributed to this, including difficulties in disambiguating affective from speech effects and perceived lack of generalizability. Here we report a framework that addresses these issues. The cornerstone of the framework is a comprehensive set of physiologically informed features, comprised of three groups: breathing depth, respiratory time quotient (RTQ), and breathing speed features. The breathing depth features capture either mental arousal or fear effects. The RTQ features capture speech production. The breathing speed features capture arousal effects due to emotional influences. The said framework appears to have broad applicability. In the naturalistic Office Tasks 2019 dataset with speaking sessions, the said features used either in regression or random forest models led to robust classification of arousal ($\overline{\text{AUC}}$ in [0.75, 0.96]) stemming from three different conditions: a) mental-emotional stressor effected through a time-pressured knowledge task; b) pure mental stressor effected through a long knowledge task; c) mental-social stressor effected through a public speech task. In the stylized CASE dataset with silent sessions, the same features and algorithms led to solid classification of arousal ($\overline{\text{AUC}}$ in [0.71, 0.85]) stemming from scary vs. non-scary movie clips.