A multi-source deep learning model for music emotion recognition

Music has been recognized as an effective tool that could be beneficial in several applications aimed at increasing people’s well-being. A personalized music recommender system can suggest playlists based on user’s preferences and considering induced emotions. Being a subjective task, it is important to define a starting solid and generalizable Music Emotion Recognition (MER) model. This model can be then refined to be adapted to the user’s specific responses, ensuring a proper interaction between the recommendation system and its user. In this paper, a MER model relying on a multi-source input, composed of songs belonging to four publicly available datasets, is presented. The proposed model is based on EfficientNetB3, designed to provide high performance while being computationally efficient. Moreover, data splitting, layer modifications, and parameter setting are proposed to reduce the model overfitting. Our proposal achieves performance comparable with those in the state of the art, providing a robust model to be adapted to a user’s emotional responses in the definition of a music recommender system.