Learning from Annotation Uncertainty: Entropy-Aware Curriculum for Speech Emotion Recognition
This project studies speech emotion recognition under annotator disagreement. Instead of collapsing labels into a single hard class, it models annotation uncertainty directly and uses entropy-aware curriculum learning to improve training.
Problem
Emotion labels in speech are often ambiguous because annotators disagree. Treating these labels as single hard targets discards useful uncertainty information and can hurt generalization.
Method
We use distribution-based supervision together with entropy-aware curriculum learning to guide the model from easier, lower-uncertainty examples toward harder, more ambiguous ones.