Music comes in many different genres and styles which are generally easy for a human listener to distinguish, but it is not so for a computer. The purpose of the research in this document is to allow a computer system to classify a sample of digital audio into one of five genres, using a mathematical/statistical pre-processing stage and an artificial neural back prop network which has been trained specifically for the task.
The selection of musical characteristics to determine in the pre-processing stage is discussed and justified, with each then being tested and evaluated for its contribution to the overall classification process.
Various network architectures and training processes are explored, along with the best use of available data examples in training and testing.
The classification task is also reduced from five genres to two genres, with the results compared and a discussion on the reasons for any differences.
The best system found in this research was capable of 67% accuracy in determining the genre of a piece of music from an audio track.
The melody is made from the numbers of Pi. 1 is C, 2 is D, and so on.
Email brendanpetty@gmail.com.