An eigenvalues analysis with entropy-per-chroma feature

A variety of audio features have been proposed for describing the music in the last decades. These features are used in various ways, for instance as inputs to different music information retrieval systems. Most, if not all, aim for capturing the essential perceptive characteristics of the audio-signal that are not present in its temporal representation. In this work we analyze entropy-per-chroma feature on a basis of eigenvalues decomposition. This feature has been successfully applied on audio-fingerprinting-based audio recognition systems. We show that involving eigenvalues decomposition contributes for an even more robust audio feature which may be used in audio matching task. In the experiments we used a set of performances of polyphonic music. We equated two versions of a same performance in order to prove that both performances yield almost the same set of features despite being performance by another musician and/or different instruments.