Speech Emotion Recognition Using Transfer Learning: A Comparative Study

Yunus Korkmaz

doi:10.1002/widm.70073

Return

Speech Emotion Recognition Using Transfer Learning: A Comparative Study

2026-02-19

PRE

Yunus Korkmaz *

DOI：10.1002/widm.70073

Original

Original request for help

Save

Abstract

En 中文

Speech emotion recognition (SER) is a critical research area at the intersection of affective computing and audio signal processing. Traditional approaches often require extensive manual feature engineering and large labeled datasets, which can limit performance in real-world applications. Recently, transfer learning with pretrained audio neural networks has gained significant traction for SER, leveraging knowledge from large-scale audio corpora to improve recognition accuracy and generalization. This review presents a comprehensive overview of transfer learning-based SER systems, with a particular focus on pretrained audio models such as YAMNet, VGGish, Wav2Vec2, and so on. Two main paradigms were examined in this study. First, feature embedding extraction, where pretrained models serve as fixed feature encoders. Second, fine-tuning strategies, where model weights are partially or fully updated on emotion-specific corpora. The article categorizes and compares state-of-the-art studies across multiple datasets, including RAVDESS, EmoDB, IEMOCAP, CREMA-D, and so on, while discussing commonly used evaluation metrics such as accuracy, precision, recall, F1 score, and AUC. Comparative tables are presented to highlight methodological trends and performance differences between feature-based and fine-tuned SER systems. This overview aims to guide researchers in selecting appropriate models and strategies for future SER tasks and to identify open challenges and future research directions in transfer learning for emotion recognition from speech signals.

Keywords:

artificial intelligence

deep learning

self supervised learning

speech emotion recognition

transfer learning

transformer