Date of Award

5-2018

Document Type

Thesis open access

Department

Computer Science

First Advisor

Matthew Hibbs

Abstract

There have been several attempts to classify music with content-based machine learning approaches. Most of these projects followed a similar procedure with a Deep Belief Network. In this project, we examined the performance of convolutional neural networks (CNN) and recurrent neural networks (RNN) as well as other components of a classification architecture, such as the choice of dataset, pre-processing techniques, and the sample size. Under a controlled environment, we discovered that the most successful architecture was a Mel-spectrogram combined with a CNN. Although our results fell behind the state-of-the-art performance, we outperform other music classification studies that use a CNN by a large margin. By performing binary classification, we also discovered individuality across genres that caused inconsistent performance.

Share

COinS