Publication

Abstract : Audio emotion recognition plays a vital role in applications such as mental health assessment, customer service automation, and interactive systems. However, manually annotating emotional audio data is often tedious, expensive, and prone to inconsistencies, hindering the development of robust machine learning models. To address these challenges, this study proposes a lightweight Deep Reinforced Active Learning (DRAL) framework that efficiently identifies the most informative and representative samples from large-scale audio datasets, thereby significantly reducing labeling requirements. The dataset is systematically divided into training, testing, seed, agent, and support sets. A Conformer model with HuBERT embeddings is employed as the base architecture, initially trained on the seed-labelled data. The framework leverages an actor–critic reinforcement learning strategy, enabling the agent to dynamically select optimal samples from the support set for labeling. Experimental evaluations on the Audio Sentiment Dataset, TESS, and CREMA-D reveal a 72–78% reduction in labelled data needs, while maintaining strong performance with accuracies of 73%, 70%, and 96%, respectively. The results demonstrate that the proposed approach not only reduces annotation and computational costs but also enhances the scalability and efficiency of audio emotion recognition systems.