Content-based Music Similarity with Triplet Networks

Our network is trained using triplets of songs such that two songs by the same In this paper, we explore the feasibility of using Triplet artist are embedded closer to one another than to networks, a variant of Siamese networks (Bromley et al., a third song by a different artist. We compare 1994), for content-based music recommendation. In this two models that are trained using different ways context, a Triplet network learns an embedding of an item of picking this third song: at random vs. based such that the item is close to other similar items and far on shared genre labels. Our experiments are conducted from dissimilar items in the embedding space. To train using songs from the Free Music Archive the network, we will consider songs by the same artist to and use standard audio features. The initial results be similar and songs by all other artists to be dissimilar.