Deploying Deep Belief Nets for content based audio music similarity

In this paper a method for computing an audio based similarity between music excerpts is presented. The method consists of three main parts, with the first step being feature extraction, which involves the calculation of three feature sets that correspond to music timbre, rhythm and harmony. Next, for each feature set a Deep Belief Network was trained without supervision on a large music collection. The respective distances of the output units of the Deep Belief Networks between two music excerpts are computed, normalized and finally combined to form the distance measure. The proposed method was evaluated on the MIREX 2013 Audio Music Similarity task. Results are encouraging, however, they indicate that the harmonic similarity component degrades the performance.