Exploring Zero-Shot Emotion Recognition in Speech Using Semantic-Embedding Prototypes

Speech Emotion Recognition (SER) makes it possible for machines to perceive affective information. Our previous research differed from conventional SER endeavours in that focused on recognising unseen emotions speech autonomously through machine learning. Such a step would enable the automatic leaning of unknown emerging emotional states. This type learning framework, however, still relied manual annotations obtain multiple samples each emotion. In order reduce this additional workload, herein, we propose zero-shot framework employing per-emotion semantic-embedding paradigm describe SER, instead using sample-wise descriptors. Aiming optimise relationship between emotions, prototypes, and samples, includes two types strategies: Sample-wise emotion-wise These strategies apply novel process respectively, via specifically designed prototypes. We verify utility these approaches by performing an extensive experimental evaluation corpora three aspects, namely influence different strategies, emotional-pair comparison, selections prototypes paralinguistic features. The results indicate is applicable use emotion recognition speech, despite choosing optimal