Automatic Cover Song Detection: Moving from High Scores to General Classification

I propose a cover song detection system based on multiple feature streams, feature normalization, and supervised classification. The system performs well not only when prior knowledge of the number of covers in a database is given, but also when a reference/test track pair is given without prior knowledge of the structure of the database. On the “covers80” test set, the proposed cover song detection system identifies 85.0% of the reference/cover pairs in the high score framework. This is a 25.9% relative improvement from the highest published score on the “covers80” test set. For the general classification task, the system identifies covers with the probability of missed detection of 13.8% and probability of false alarm at 2.7%.