A fast audio similarity retrieval method for millions of music tracks

We present a filter-and-refine method to speed up nearest neighbor searches with the Kullback–Leibler divergence for multivariate Gaussians. This combination of features and similarity estimation is of special interest in the field of automatic music recommendation as it is widely used to compute music similarity. However, the non-vectorial features and a non-metric divergence make using it with large corpora difficult, as standard indexing algorithms can not be used. This paper proposes a method for fast nearest neighbor retrieval in large databases which relies on the above approach. In its core the method rescales the divergence and uses a modified FastMap implementation to speed up nearest-neighbor queries. Overall the method accelerates the search for similar music pieces by a factor of 10–30 and yields high recall values of 95–99% compared to a standard linear search.