Automated Thematic Composer Classification Using Segment Retrieval
Jacob Edward Galajda, Kien A. Hua · 2024
Music is a collection of measures expertly crafted and placed together by composers. However, most composer classification models aim to recognize entire songs and may not be suited for thematic analysis that analyzes each collection of measures within a song. Additionally, composer classification techniques rely on large amounts of songs, and previously presented transformer neural network iterations cannot accommodate samples outside of the training dataset due to their tokenization process. In this paper, we propose a lightweight retrieval technique that achieves the single composer benchmark accuracies presented by previous classification models at a fraction of the computing resources. This solution can be applied to variable-length MIDI songs from composers both included and excluded in the training dataset, and it achieves 100% accuracy in our performance study.