Deep Learning Representations for Music Similarity

Summary

The rapid growth of digital technology over the past few decades, especially with the rise of music streaming platforms and online music services, has led to a huge increase in available music data. This creates both new opportunities and challenges in managing and finding relevant content efficiently. This thesis explores how to extract meaningful information from deep learning models and combine these features with handcrafted audio features to improve song similarity retrieval, while also analyzing the impact of distance metrics and normalization methods in optimizing the accuracy and performance of the similarity retrieval process. The results of this research contribute to advances in the field of Music Information Retrieval (MIR) and can potentially be useful for tasks such as song recommendations and playlist creation.