Correlating Extracted and Ground-Truth Harmonic Data in Music Retrieval Tasks

We show that traditional music information retrieval tasks with well-chosen parameters perform similarly using computationally extracted chord annotations and groundtruth annotations. Using a collection of Billboard songs with provided ground-truth chord labels, we use established chord identification algorithms to produce a corresponding extracted chord label dataset. We implement methods to compare chord progressions between two songs on the basis of their optimal local alignment scores. We create a set of chord progression comparison parameters defined by chord distance metrics, gap costs, and normalization measures and run a black-box global optimization algorithm to stochastically search for the best parameter set to maximize the rank correlation for two harmonic retrieval tasks across the ground-truth and extracted chord Billboard datasets. The first task evaluates chord progression similarity between all pairwise combinations of songs, separately ranks results for ground-truth and extracted chord labels, and returns a rank correlation coefficient. The second task queries the set of songs with fabricated chord progressions, ranks each query’s results across ground-truth and extracted chord labels, and returns rank correlations. The end results suggest that practical retrieval systems can be constructed to work effectively without the guide of human ground-truthing.