Music Genre Classification Using Convolutional Neural Network

Abstract

Feature extraction is essential in Music
Information Retrieval (MIR), but traditional
methods like MFCC are often ineffective for
music genre classification. This study
introduces an algorithm that utilizes
spectrograms and Convolutional Neural
Networks (CNNs) for better classification
performance. Unlike MFCC, spectrograms
capture detailed musical features like pitch
and flux. Our approach uses a CNN to
generate four feature maps from the
spectrogram, which reflect trends over time
and frequency. A subsampling layer reduces
dimensions and improves resistance to pitch
and tempo changes. The multi-layer
perceptron (MLP) classifier achieves a
72.4% accuracy on the Tzanetakis dataset,
surpassing MFCC's performance.