Human non-linguistic emotional vocalizations
Andrey Anikin · 2015
Despite the recent upsurge of interest in emotional non-speech sounds, very little naturalistic material is available. In this project a corpus of non-acted emotional vocalizations from Youtube videos is introduced, analyzed acoustically and tested experimentally. A rating experiment combined with acoustic analysis in Study 1 confirms that these sounds can be correctly classified by both human raters and statistical models, albeit with lower hit rates than in studies of play-acted vocalizations. Interestingly, the overall accuracy of recognizing emotion in the predominantly English-language corpus is similar for Russian-, Swedish- and English-speaking subjects, with no in-group advantage. Study 2 compares naturalistic Youtube vocalizations with play-acted vocalizations from animated cartoons and two previously published corpora: by Lima et al. (2013) and Belin et al. (2008). Based on acoustic analysis and statistical modeling, it appears that some emotions are expressed very similarly in real life and in the studio (amusement, disgust, pleasure), whereas others show systematic differences (fear, anger, sadness). To address the difficulties related to the choice of emotional categories and engage with dimensional models of emotion, a label-free triad task is then used in Study 3 to explore the structure of acoustic-emotional space and the nature of distinguished categories. It suggests a two- or three-dimensional structure broadly compatible with the circumplex model, but with certain peculiarities related to the acoustic nature of the stimuli. The tentative conclusion is that many of the real-life non-linguistic emotional vocalizations explored in this study may best be described in terms of a few flexibly used and graded call types, rather than in terms of strictly emotion-specific acoustic patterns. (Less)