Music Genre Classification Based on Chroma Features and Deep Learning

Abstract

Music genre classification is an important branch of content-based music signal analysis. It is a challenging task in the field of music information retrieval (MIR). At present, the method based on deep learning has achieved good results. This paper constructs a neural network framework for music genre classification based on chroma feature. The chroma feature can represent the time domain and the frequency domain of music character and consider the existence of harmony. Besides, it is independent of the timbre, volume, absolute pitch, which are completely irrelevant to the genre classification. It is relatively robust to the background noise and can represent the primary information such as monophonic and polyphonic music distribution. In this paper, we estimate the type of music audio based on chroma feature combined with deep learning network. We input this feature into VGG16 network for training, and improve the last three layers. In the experiment, the classifier is trained by GTZAN dataset. The experimental results show that the framework can obtain higher classification accuracy and better performance.