Deep Metric Learning with Music Data

Music Information Retrieval has often relied on supervised learning with handpicked features. These supervised models are often trained with a fixed set of labels and require to be retrained whenever new labels are introduced to the dataset. In this paper we develop a distance metric between a database of songs provided to it and use it to mine deeper and more complex relationships between songs that may not be easily quantified. We achieve this by the usage of Deep Metric Learning using a Siamese network. We have developed a metric between multiple songs of artist. Our developed model is successful in maintaining a healthy separation between artists and able to also encode deeper relations such as genre. The model performs well even when new labels are introduced to the dataset and boasts comparable accuracy to previous research done in this field as it is well generalized due to hard triplet mining and can be used to handle multiple tasks like artist prediction, genre auto-tagging and content-based retrieval.