Deep learning methods for cover song identification
View/
Keywords
Deep learning ; Music Information Retrieval ; CNN ; CSIAbstract
Cover song identification (CSI) is the task of determining whether a given recording of a song is a new performance other than the original version. Automatically detecting cover versions has plenty of applications in the music industry as well as copyright law. In this Thesis we present a methodology for CSI based on Convolutional Neural Networks (CNN) and Metric Learning. The model is trained on medium-size datasets of cover songs using a variation of the Triplet Loss, called Angular Loss. The experiments showcase the performance of the proposed CNN model on English and Greek sets of cover songs, as well as other approaches based on deep learning. Our findings demonstrate that the proposed method exhibits viable performance for the specific use case, achieving high scores on the classification and ranking tasks. This, along with the fact that the model can run with minimal hardware requirements, make our method an ideal candidate for real-world applications. To further illustrate this point, we designed a proof of concept of such a system. Finally, as part of this Thesis, we created two new open-source datasets for CSI, that can be used for training or evaluation.