Fast and Scalable System for Automatic Artist Identification

Digital music technologies enable users to create and use large collections of music. One of the desirable features for users is the ability to automatically organize the collection and search in it. One of the operations that they need is automatic identification of tracks' artists. This operation can be used to automatically classify new added tracks to a collection. Additionally, the user can use this operation to identify the artist of an unknown track. The artist name of a track can help the user find similar music. In this paper, we introduce a fast and scalable system that can automatically identify the artist of music tracks. This system is creating a signature for each track that is a compact representation of the tracks. The tracks' signatures of an artist are then used to create a signature for that artist. A similarity measure is also defined to measure the distance or dissimilarity of two signatures. This similarity measure is based on graph matching. To identify the signature of an unknown track, the signature of that track is compared with the artists' signatures and the nearest artist is selected as the artist of the track. The accuracy of the system on the artist20 dataset is 71.5% which is better than previously reported results on this dataset. In comparison to other proposed methods, the artist model creation and model updates are faster and more scalable in our system. Additionally, the search time of our system is less than other systems and only depends on the number of artist.