Speech emotion recognition (SER) systems are typically trained on datasets labeled with a single ground-truth emotion category. But that category can refer to what the speaker intended to convey or what listeners actually perceive, and these are not always the same. A new arXiv paper, "Label Authority in Speech Emotion Recognition: Intended and Perceived Ground Truth in Emozionalmente," takes up this problem directly.
The authors argue that an acted emotion and the emotion heard by listeners describe different aspects of the same recording. They examine this distinction using the Emozionalmente dataset, which appears designed to separate intended and perceived labels. The work questions whether a single authoritative label is sufficient for training robust SER models.
If intended and perceived emotions diverge systematically, then models trained on one type of label may not generalize to real-world listening conditions. The study's significance lies in pushing the field to reconsider what "ground truth" means in emotion recognition, and whether both perspectives should be captured explicitly in datasets and evaluation metrics.