Emotion Classification from Speech by an Ensemble Strategy

Abstract

Humans are prepared to comprehend each other's emotions through subtle body movements and speech expressions, and from those, they change the way they deliver/understand messages when communicating between them. Socially assistive robots need to empower their ability in recognizing emotions in a way to change the interaction with humans, especially with elders. This paper presents a framework for speech emotion prediction supported by an ensemble of distinct out-of-the-box methods, being the main contribution of the integration of the outputs of those methods in a single prediction consistent with the expression presented by the system's user. Results show a classification accuracy of 75.56% over the RAVDESS dataset and 86.43% in a group of datasets constituted by RAVDESS, SAVEE, and TESS.