No abstract

This paper presents a novel dataset of songs by non-superstar artists in which a set of musical data is collected, identifying for each song its musical structure, and the emotional perception of the artist through a categorical emotional labeling process. The generation of this preliminary dataset is motivated by the existence of biases that have been detected in the analysis of the most used datasets in the field of emotion-based music recommendation. This new dataset contains 234 min of audio and 60 complete and labeled songs. In addition, an emotional analysis is carried out based on the representation of dynamic emotional perception through a time-series approach, in which the similarity values generated by the dynamic time warping (DTW) algorithm are analyzed and then used to implement a clustering process with the K-means algorithm. In the same way, clustering is also implemented with a Uniform Manifold Approximation and Projection (UMAP) technique, which is a manifold learning and dimension reduction algorithm. The algorithm HDBSCAN is applied for determining the optimal number of clusters. The results obtained from the different clustering strategies are compared and, in a preliminary analysis, a significant consistency is found between them. With the findings and experimental results obtained, a discussion is presented highlighting the importance of working with complete songs, preferably with a well-defined musical structure, considering the emotional variation that characterizes a song during the listening experience, in which the intensity of the emotion usually changes between verse, bridge, and chorus.

No abstract