Querying Large Collections of Music for Similarity

We present a system capable of performing similarity queries against a large archive of digital music. Users are able to search for songs which “sound similar” to a given query song, thereby aiding the navigation and discovery of new music in such an archive. Our technique is based on reduction of the music data to a feature space of relatively small dimensionality (1248 feature dimensions per song); this is accomplished using a set of feature extractors which derive frequency, amplitude, and tempo data from the encoded music data. Queries are then performed using a k-nearest neighbor search in the feature space. Our system allows subsets of the feature space to be selected on a per-query basis. We have integrated the music query engine into an online MP3 music archive consisting of over 7000 songs. We present an evaluation of our feature extraction and query results against this archive.