Music genre classification using deep neural networks and data augmentation

Abstract

Music is an indispensable part of spiritual life. Today, humanity’s musical treasure is truly huge and precious, and the number of musical works is constantly increasing. Computers, machine learning, and deep learning have greatly aided in the storing, organizing, searching, and enjoying of musical works in priceless treasures. Many music databases have been built for such music-data processing studies. One operation that needs to be handled automatically for musical works is musical genre classification (MGC). This paper presents new research results on MGC for GTZAN music data. Deep neural networks, such as convolutional neural networks (CNN), long short-term memory (LSTM), gated recurrent units (GRU), and capsule neural networks (CSN), have produced excellent results when combined with data augmentation methods such as splitting audio files, noise addition, and pitch shifting. A classification accuracy of 99.91% for the ten musical genres of GTZAN was achieved using the CSN model with the Mel spectrogram as input features and data enhanced by the aforementioned methods. This classification accuracy outperformed that of all previous GTZAN classification accuracy studies.