Melodic Similarity among Incipits; a Deep Learning Approach

Abstract

This project introduces transformer models to the field of melodic similarity. Originating from the Natural Language Processing field; transformer models use sequential data to find relationships between individual parts of a sequence. Melodies can be regarded as sequences of notes, making melodies viable as input for transformer models. Deep learning approaches like RNNs have already proven to compete with the state-of-the-art alignment algorithm regarding melodic similarity. This project aims to discover whether a transformer model is capable of finding groups of similar melodies in the Meertens Tune Collection(MTC). Using triplet loss, transformers are capable of training an embedding space with the goal to cluster similar melodies. Here we show that transformers can reach a MAP score of 0.36 and P@1 score of 0.48 on unseen data from the MTC dataset. We also made a comparison between different input forms, where using whole melodies, random triplets and a complex feature set is preferable over other variations of the input. We discovered that the tokenisation process put restrictions on the feature selection process, which negatively impacted model performance. Although not competing with the state-of-the-art methods, transformers show potential in solving problems related to melodic similarity.