AMRD: Adaptive Multi-Teacher Relational Distillation for Lightweight Speech Emotion Recognition

Researchers introduced Adaptive Multi-teacher Relational Distillation (AMRD) for lightweight speech emotion recognition.

AMRD adjusts teacher weights batch by batch using logit-similarity coherence and adds a relational loss to preserve inter-sample structure.

On IEMOCAP and CREMA-D, it outperformed single-teacher baselines across multiple student architectures.

The approach makes real-time edge and mobile speech emotion recognition more practical by improving small-model accuracy.