Music genre classification based on res-gated CNN and attention mechanism

• Published in: Multimedia tools and applications Vol. 83; no. 5; pp. 13527 - 13542
• Main Authors: Xie, Changjiang, Song, Huazhu, Zhu, Hao, Mi, Kaituo, Li, Zhouhan, Zhang, Yi, Cheng, Jiawen, Zhou, Honglin, Li, Renjie, Cai, Haofeng
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.02.2024; Springer Nature B.V
• Subjects: Artificial neural networks; Audio data; Classification; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Feature extraction; Genre; Multimedia; Multimedia Information Systems; Music; Special Purpose and Application-Based Systems; Transformers; Music genre classification; CNN; Transformer
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-023-15277-1

The amount of digital music available on the internet has grown significantly with the rapid development of digital multimedia technology. Managing these massive music resources is a thorny problem that powerful music media platforms need to face where music genre classification plays an important role, and a good music genre classifier is indispensable for the research and application of music resources in the related aspects, such as efficient organization, retrieval, recommendation, etc. Due to convolutional networks’ powerful feature extraction capability, more and more researchers are devoting their efforts to music genre classification models based on convolutional neural networks (CNNs). However, many models do not combine the musical signal features for effective design of the convolutional structure, which cause a simpler convolutional network part of the model and weaker local feature extraction ability. To solve the above problem, our group proposes a model using a 1D res-gated CNN to extract local information of audio sequences rather than the traditional CNN architecture. Meanwhile, to aggregate the global information of audio feature sequences, our group applies the Transformer to the music genre classification model and modifies the decoder structure of the Transformer according to the task. The experiments utilize the benchmark datasets, including GTZAN and Extended Ballroom. Our group conducted contrastive experiments to verify our model, and experimental results demonstrated that our model outperforms most of the previous approaches and can improve the performance of music genre classification.