Key-independent classification of harmonic change in musical audio

We introduce a novel method for describing the harmonic development of a musical signal by using only low-level audio features. Our approach uses Euclidean and phase distances in a Tonal Centroid space. Both measurements are taken between successive chroma partitions of a harmonically segmented signal, for each of three harmonic circles representing fifths, major thirds, and minor thirds. The resulting feature vector can be used to quantify a string of successive chord changes according to changes in chord quality and movement of the chordal root. We demonstrate that our feature set can provide both unique classification and accurate identification of harmonic changes, while resisting variations in orchestration and key.