Parallel Deep Neural Networks for Musical Genre Classification: A Case Study

Abstract

Musical genres are labels created to categorize the universe of music. A music genre is characterized by its unique form or style, including instrumentation, rhythmic structure, and harmonic content. It is conventional for a large collection of music to be structured using genre hierarchies. Automatic music genre classification is gaining attention in recent years due to the large amount of available digital music on the web and the latest advances in artificial intelligence. In particular, various deep learning-based approaches have delivered promising results in this domain. In this paper, we present a case study of the PRCNN framework (2017), which parallelizes CNN and bi-directional GRU to capture both spatial and temporal signals from music spectrograms. In our study, we designed our model based on the proposed concept but with a different model structure. Furthermore, we trained and evaluated our model using a more comprehensive dataset (FMA) with 8,252 pieces of music and 17 genres. We further validated our model on a curated dataset of 15 songs. Our model achieves an overall accuracy of 88% on the FMA dataset with above 90% accuracies in four genre categories. For the curated dataset, the model correctly classified 11 out of the 15 songs. Our experimental results provide convincing support for utilizing parallelized deep neural networks to model the concurrent spatial and temporal characteristics of music data.