Evaluation of content-based acoustic features for musical genre classification

In this thesis, we evaluate content-based acoustic features for musical genre classification. Effectiveness of various acoustic features are compared using a k-nearest neighbor (KNN) classifier. By utilizing the combinations of acoustic features, an average classification accuracy of $89\%$ for GTZAN database is achieved, which is comparable to prior work. A statistical test, McNemar's test, is applied to support the idea that musical genre is intrinsically related to content-based acoustic features. Especially for some genres, we are able to identify the particular associated acoustic property. In addition, by comparing our KNN results to a psychoacoustic listening experiment, we associate various human perceptual dimensions with low-level acoustic features