Novel mid-level audio features for music similarity

Large-scale systems for automatic content-based music recommendation require efficient computation of signal descriptors that are robust and relevant with regard to human perception in order to process extensive music archives. In this publication, a set of mid-level audio features suitable for efficient characterization of musical signals with regard to automatic music similarity estimation is proposed. These descriptors are dedicated to model timbre modulation, song dynamics, rhythmic qualities and melodic properties of a music piece in a semantic context. An outline of implementation details and peculiarities will be given. Based on a well defined benchmark criterion for music similarity assessment the strengths and weaknesses of the introduced features will be depicted and discussed accordingly.