Theory and Evaluation of a Bayesian Music Structure Extractor

We introduce a new model for extracting end points of music structure segments, such as intro, verse, chorus, break and so forth, from recorded music. Our methods are applied to the problem of grouping audio features into continuous structural segments with start and end times corresponding as closely as possible to a ground truth of independent human structure judgements. Our work extends previous work on automatic summarization and structure extraction by providing a model for segment end-points posed in a Bayesian framework. Methods to infer parameters to the model using Expectation Maximization and Maximum Likelihood methods are discussed. The model identifies all the segments in a song, not just the chorus or longest segment. We discuss the theory and implementation of the model and evaluate the model in an automatic structure segmentation experiment against a ground truth of human judgements. Our results shows a segment boundary intersection rate break-even point of approximately 80%.