A Filter-and-Refine Indexing Method for Fast Similarity Search in Millions of Music Tracks

We present a filter-and-refine method to speed up acoustic audio similarity queries which use the Kullback-Leibler divergence as similarity measure. The proposed method rescales the divergence and uses a modified FastMap [1] implementation to accelerate nearest-neighbor queries. The search for similar music pieces is accelerated by a factor of 10−30 compared to a linear scan but still offers high recall values (relative to a linear scan) of 95− 99%. We show how the proposed method can be used to query several million songs for their acoustic neighbors very fast while producing almost the same results that a linear scan over the whole database would return. We present a working prototype implementation which is able to process similarity queries on a 2.5 million songs collection in about half a second on a standard CPU.