Department of Computer Engineering2024-11-092019978-1-5386-9490-92156-2318N/A2-s2.0-85073066888https://hdl.handle.net/20.500.14288/12463Dyadic interactions are reflective of mutual engagement between their participants through different verbal and non-verbal voicing cues. This study aims to investigate the effect of these cues on continuous emotion recognition (CER) using speech and head motion data. We exploit the non-verbal vocalizations that are extracted from speech as a complementary source of information and investigate their effect for the CER problem using gaussian mixture and convolutional neural network based regression frameworks. Our methods are evaluated on the CreativeIT database, which consists of speech and full-body motion capture under dyadic interaction settings. Head motion, acoustic features of speech and histograms of non-verbal vocalizations are employed to estimate activation, valence and dominance attributes for the CER problem. Our experimental evaluations indicate a strong improvement of CER performance, especially of the activation attribute, with the use of non-verbal vocalization cues of speech.EngineeringIndustrial engineeringEngineeringElectrical electronic engineeringsConference proceeding5559526000803001