Melodic Similarity: A Re-examination of the MIREX2005 Data

Despite a considerable body of research, there is no clarity about the basic properties of melodic similarity, such as whether or not it constitutes a metric space, or whether it is a more complex phenomenon. An experiment conducted by Typke et al., used as a basis for the MIREX2005 melodic-similarity modelling contest, represents a particularly rich source of data. In the experiment, for each of eleven queries (melodies taken from RISM A/II), about 25 experts ranked some of about 50 candidates for similarity with the query. A ‘Monte Carlo’ approach has been taken in re-examining this data, simulating data in the same form on the basis of simple assumptions about the nature of melodic similarity. Statistical properties of the actual data were compared with the same properties for 10,000 sets of simulated data, allowing estimation of the significance of differences found. In terms of overall measures such as the ranking profile for each candidate, quite good simulations (i.e., sets of simulated data in which the original falls within the second and third quartiles in the measured property) arose from stochastic ranking based only on the mean and variance of the actual ranking for each candidate and on the likelihood of the candidate being selected for ranking. However, the simulations did show evidence, in a substantial minority of cases, of an effect for some candidates to be ranked higher or lower dependent on the presence of another candidate, and of the influence of similarity between candidates.