Music structure segmentation algorithm evaluation: Expanding on MIREX 2010 analyses and datasets

Music audio structure segmentation has been a task in the Music Information Retrieval Evaluation eXchange (MIREX) since 2009. In 2010, five algorithms were evaluated against two datasets (297 and 100 songs) with an almost exclusive focus on western popular music. A new annotated dataset significantly larger in size and with a more diverse range of musical styles became available in 2011. This new dataset comprises over 1,300 songs spanning pop, jazz, classical, and world music styles. The algorithms from the 2010 iteration of MIREX are re-evaluated against this new dataset. This paper presents a detailed analysis of these evaluation results in order to gain a better understanding of the current state-of-the-art in automatic structure segmentation. These expanded analyses focus on the interaction of algorithm performance and rankings with datasets, musical styles, and annotation level. Because the new dataset contains multiple annotations for each song, we also introduce a baseline for expected human performance for this task.