In-place versus re-build versus re-merge: Index maintenance strategies for text retrieval systems

Attempts to categorise music by extracting audio features from a sample have had mixed results. Some categories such as classical are easy to identify but attempts to distinguish between various types of popular music yield poor results. Part of the difficulty is that humans also disagree with each other when classifying music. We report on experiments that compare human classification of music samples to that based on audio feature extraction and machine learning techniques. We extracted a set of audio features and applied a range of machine learning techniques to a set of 128 pieces of music. Our work demonstrates that a single feature and a simple machine learning approach achieve results that are almost as consistent as humans for the same task. Further experiments revealed an even greater inconsistency amongst humans in selecting categories for music. Using a self-organising map on the same set of pieces and features produced some meaningful song clusters, that is, pieces by the same artist or composer, or of the same genre, were grouped together. It also showed some of the same cross-genre relationships shown by the human-based classifications