SPBKD: Subjective Perceptual Bias Knowledge Distillation Through Attention, Masking and Prompt Tuning for Speech Emotion Recognition

View at publisher Summarise with VIRAT
Authors: Wei Zhang, Chunyang Tian
Cited by: 0
References: 54

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Speech Emotion Recognition (SER) plays a pivotal role in human-computer interaction. Conventional SER models typically rely on acoustic features. However, these features often exhibit weak robustness due to the inherent ambiguity and subjectivity in emotional expression and perception. Furthermore, relying solely on consensus labels discards rich evaluative information available in datasets. To address these challenges, we propose a Subjective Perceptual Bias Knowledge Distillation (SPBKD) framework. This method fully exploits the diverse evaluation information in the IEMOCAP and MSP-Improv datasets, where multiple raters provide subjective categorical annotations for the same sample. By encoding these perceptual differences, we enable the teacher model to achieve more accurate emotion recognition. Specifically, we design a knowledge distill-ation framework where the teacher model incorporates both acoustic features and encoded rater subjective perceptual bias features. The student model then learns from the soft labels generated by the teacher, thereby acquiring better generalization capabilities for the inference phase. Experimental results demonstrate that on the IEMOCAP dataset, SPBKD achieves 79.4% WA and 80.3% UA, representing improvements of 2.3% over the baseline. Similarly, on the MSP-Improv dataset, it achieves 75.9% WA and 74.1% UA, outperforming the baseline by 0.8% and 1.3%, respectively.

Fields of Study & Subjects

Emotion and Mood Recognition Speech Recognition and Synthesis Speech and Audio Processing

Publication details

Journal / source
IEEE Transactions on Audio Speech and Language Processing
Volume
34
Pages
2140–2153
Publication date
1 January 2026
Language
EN