Affective Computing in Music: A Critical Review of Deep Learning’s Role and Its Alignment with Psychological Emotion Models

The precise identification of musical emotion holds substantial academic and practical value, with critical applications in music recommendation, therapeutic interventions, and real-time human–computer interaction. Conventional paradigms for music emotion recognition, predominantly based on manual feature engineering and conventional machine learning, exhibit major impediments, including pronounced label subjectivity and insufficient generalizability, which have collectively hampered the further development of this domain. The development of convolutional neural networks (CNNs), recurrent neural networks (RNNs), and graph neural networks (GNNs) has expanded the range of automated approaches available for music emotion recognition. Every approach shows some intrinsic advantages and weakness; thus, they have had some success but are not yet fully qualified for application requirements. This paper will systematically summarize the application of deep learning in music emotion recognition in recent years, including theoretical frameworks, technical implementations, challenges, and future research trends. Typically, we will systematically discuss the current research status of deep learning and machine learning in music emotion recognition via a bibliometric analysis and then reveal the working mechanism of deep learning, which is underestimated in most previous studies. This work would shape the development of this research field, especially for the applications of deep learning in music emotion recognition.