A Cross-Validated Study of Modelling Strategies for Automatic Chord Recognition in Audio.
John Burgoyne, Laurent Pugin, Corey Kereliuk, Ichiro Fujinaga · 2007
Although automatic chord recognition has generated a number of recent papers in MIR, nobody to date has done a proper cross validation of their recognition results. Cross validation is the most common way to establish baseline standards and make comparisons, e.g., for MIREX competitions, but a lack of labelled aligned training data has rendered it impractical. In this paper, we present a comparison of several modelling strategies for chord recognition, hiddenMarkov models (HMMs) and conditional random fields (CRFs), on a new set of aligned ground truth for the Beatles data set of Sheh and Ellis (2003). Consistent with previous work, our models use pitch class profile (PCP) vectors for audio modelling. Our results show improvement over previous literature, provide precise estimates of the performance of both old and new approaches to the problem, and suggest several avenues for future work.