Music Feature Classification: Using Techniques of Machine Learning to Detect Genre

Genre is a fluid descriptor used to categorize and classify musical works. Although it has historically been assigned manually, machine learning algorithms such as neural networks and random forests are powerful tools that can automate this process. Past research has demonstrated that training a neural network on spectral features extracted from the time series representation of a music file yields high accuracy. In this project, we studied and implemented two machine learning algorithms, the convolutional neural network and the random forest. Each method was applied to a novel dataset we extracted from 100 music files spanning 4 musical genres. In addition to using a list of traditional sound features to classify the music clips, we discussed the effectiveness of two musical attributes that are essential in music theory, time signature and key signature. We have shown using a two-layered neural network and a random forest that incorporating these two musical attributes of a piece does not increase the accuracy of genre prediction, although these two features are reasonably important when humans sort and classify music files. We also applied the convolutional neural network to the GTZAN dataset, which contains 10 music genres and 1000 music files, and obtained an accuracy of 92.42 %. This result verifies currently successful research on the application of the convolutional neural network to music genre classification