Werewolf-XL

Existing databases usually record posed or induced human behavior in individual or dyadic settings with biased annotations in which a basic emotional class label or a Valence–Arousal pair value represents the emotional states. To address this need,We created the “Werewolf-XL” database, which contains a total of 987 h of spontaneous audio-visual recordings of 129 subjects in a group interaction of nine individuals playing a conversational role-playing game called Werewolf. We provide 133,461 individual utterance-level video clips with internal (18 non-prototypical emotional categories) and external (pleasure, arousal, and dominance) emotion annotations including 14,635 samples from speaking players. Besides, the results of annotation agreement analysis show fair reliability and validity. Furthermore, we provided extensive benchmarks of unimodal and multimodal emotional recognition results. The database (the video clips, annotations, features and protocols) is made publicly available, which provides great opportunities for researchers in affective computing and beyond.