Algorithmic prediction of inter-song similarity in Western popular music

We investigate a method for automatic extraction of inter-song similarity for songs selected from several genres of Western popular music. The specific purpose of this approach is to evaluate the predictive power of different feature extraction sets based on human perception of music similarity and to develop an algorithm able to reproduce and predict human ratings. The algorithm is a linear model that was trained and tested using perceptual data. We use publicly available algorithms to extract acoustic feature values from 78 songs used in a previous perceptual experiment. Feature value differences between songs are used in a multivariate linear regression calculation to find the optimal weighting coefficients for the feature values to best approximate the human similarity perception data. We use two evaluation methods: metrical and ordinal. We use a bootstrapping approach by randomly separating the experimental data into training and testing sets. We compare the performance of this model against the G1C model by Pampalk, winner of the MIREX 2006 competition on music similarity prediction. Both models produce a rather low performance on the metrical evaluation. However, on the ordinal evaluation, the linear regression model shows encouraging results (significantly outperforming the G1C algorithm): in the triadic comparison task, it can correctly predict 52.3 ± 0.5% of the most similar pairs, while the estimated theoretical maximum, based on participant consistency on the most similar pair rankings is 78 ± 8%. In a comparison of feature sets, we found the MIR toolbox to produce the best performance