A Music Recommendation System Based on Semantic Audio Segments Similarity

In this paper we propose a novel approach for contentbased music recommendation. The main innovation of the proposed technique consists of a similarity function that, instead of considering entire songs or their thumbnail representations, analyzes audio similarities between semantic segments from different audio tracks. The rationale of our idea is that a song similarity and recommendation technique, to be more meaningful to the user from a semantic point of view, may evaluate and exploit similarities on semantic units between audio tracks. Our similarity algorithm consists of two main stages: the first step performs segmentation of the song in semantic parts. The latter assigns a similarity and recommendation score to a pair of songs, by computing the distance between the representations of their segments. To assign the global similarity and recommendation score, we consider a consistent subset of all the inter-segment distances. By adopting a graph-bases framework, we propose a graph-reduction algorithm on weighted edges that connect segments of different songs to optimize the similarity score with respect to our recommendation goal. Experiments conducted on a database of 200 audio tracks of various authors and genres show promising results.