Combining Reduction and Dense Blocks for Music Genre Classification

• Published in: Neural Information Processing Vol. 1517; pp. 752 - 760
• Main Authors: El Achkar, Charbel, Couturier, Raphaël, Atéchian, Talar, Makhoul, Abdallah
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2021; Springer International Publishing
• Series: Communications in Computer and Information Science
• Subjects: CNN; Music genre classification; Music recommendation systems; STFT spectrogram
• ISBN: 9783030923099; 3030923096
• ISSN: 1865-0929; 1865-0937
• DOI: 10.1007/978-3-030-92310-5_87

Embedding music genre classifiers in music recommendation systems offers a satisfying user experience. It predicts music tracks depending on the user’s taste in music. In this paper, we propose a preprocessing approach for generating STFT spectrograms and upgrades to a CNN-based music classifier named Bottom-up Broadcast Neural Network (BBNN). These upgrades concern the expansion of the number of inception and dense blocks, as well as the enhancement of the inception block through reduction block implementation. The proposed approach is able to outperform state-of-the-art music genre classifiers in terms of accuracy scores. It achieves an accuracy of 97.51% and 74.39% over the GTZAN and the FMA dataset respectively. Code is available at https://github.com/elachkarcharbel/music-genre-classifier.