SOUNDSCOUT: A SONG RECOMMENDER BASED ON SOUND SIMILARITY FOR HUGE COMMERCIAL MUSIC ARCHIVES

We present Soundscout, a song recommender operating on huge commercial music archives. In contrast to many playlist generation tools which support the consumer in handling local collections, Soundscout focuses on efficient metadata generation from large bodies of music files. The system automatically generates similarity relations between songs based on acoustic features extracted from the sound files and combines the thus achieved results with a variety of preference models as filters to minimize the risk of inappropriate recommendations. In the paper we describe how we significantly improved the applicability of state-of-the-art MIR techniques by introducing an efficient k-nearest neighbour algorithm feasible for handling a huge number of tracks. Moreover we introduce filter mechanisms and presentation strategies for improving the acceptability of the sounding similar recommendations. Last but not least we present a Web application of Soundscout for conducting user acceptance tests, and present results from a first exploratory study.