BEES: A New Acoustic Task for Blended Emotion Estimation in Speech

Acoustic emotion recognition (AER) is a challenging yet crucial area in affective computing. Despite the clear importance of perceiving vocal emotion blends with varying intensities for understanding the emotional complexity of everyday life, most existing studies remain limited to single-label emotion classification tasks. This paper introduces a novel acoustic task for blended emotion estimation in speech (BEES), which aims to capture the intricate emotional distributions of co-existing emotions with varied intensities in voices. The BEES task stands out by defining a precise reference for the emotion intensity of vocal blends, introducing innovative evaluation metrics and providing robust baseline results. Additionally, our study highlights the significant impact of self-supervised learning representations, voice naturalness, and gender on BEES performance. By addressing the limitations of traditional AER methods and offering a comprehensive framework for analyzing blended emotions, this work paves the way for advanced computational AER systems capable of handling the complexity of vocal emotion blends.