Music Genre Classification Using Convolutional Neural Network Method

Music genre classification is a prominent challenge in music information retrieval research. Most techniques for music genre classification employ machine learning algorithms to classify feature vectors extracted from sound recording segments. Several datasets have been utilized in experiments to evaluate classification accuracy in this domain. However, the datasets commonly used are relatively small, and much of the research on music genre classification continues to focus on these limited datasets. This research utilizes data from three main sources: (1) metadata derived from the Million Song Dataset, (2) the CD2C tagtraum genre annotations dataset as ground truth for genre labels, and (3) the Million Song Dataset mapping archive, which provides metadata mapping to SpotifyAPI, along with (4) acoustic features extracted via SpotifyAPI. The resulting dataset comprises 64,088 music tracks spanning five popular genres: Rock, Electronic, Jazz, Pop, and Rap. The objective of this research is to develop a model capable of predicting music genres based on acoustic features extracted through SpotifyAPI. To achieve this, a Convolutional Neural Network (CNN) architecture was implemented, with a focus on hyperparameter optimization, particularly in terms of the number of layers and the type of weight initialization. The optimal hyperparameters were identified as a two-layer CNN architecture utilizing Xavier Normal Initialization for weight initialization. The resulting model achieved a weighted F1-score of 0.7030 during testing.