AMRD: Adaptive Multi-Teacher Distillation for Speech Emotion Recognition
July 27, 2026
AMRD is a knowledge distillation method for lightweight speech emotion recognition on edge devices. It uses a one-class SVM to weight teachers based on batch coherence and applies a relational distillation loss to capture inter-sample similarity structures.
HOW THIS AFFECTS YOU
●
builderThis can help you deploy high-quality emotion recognition models on resource-constrained edge hardware.
●
researcherThe inclusion of relational distillation loss provides a more holistic way to compress teacher-student models.