Learning a Feature Space for Similarity in World Music

In this study we investigate computational methods for assessing music similarity in world music. We use state-ofthe-art audio features to describe musical content in world music recordings. Our music collection is a subset of the Smithsonian Folkways Recordings with audio examples from 31 countries from around the world. Using supervised and unsupervised dimensionality reduction techniques we learn feature representations for music similarity. We evaluate how well music styles separate in this learned space with a classification experiment. We obtained moderate performance classifying the recordings by country. Analysis of misclassifications revealed cases of geographical or cultural proximity. We further evaluate the learned space by detecting outliers, i.e. identifying recordings that stand out in the collection. We use a data mining technique based on Mahalanobis distances to detect outliers and perform a listening experiment in the ‘odd one out’ style to evaluate our findings. We are able to detect, amongst others, recordings of non-musical content as outliers as well as music with distinct timbral and harmonic content. The listening experiment reveals moderate agreement between subjects’ ratings and our outlier estimation.