Classifying Classical Music Genres with Neural Networks

Current neural network models can process and interpret music for tasks such as melody completion and genre or style classification. However, previous classification tasks do not account specifically for distinct composition styles of different classical music periods, often focusing instead on modern genres. To bridge this gap, this project investigates the use of natural language processing techniques to classify musical excerpts from the Baroque, Classical, and Romantic periods. A curated dataset of samples representative of the three eras, converted to the OctupleMIDI format, was used to train a Sentence Transformers model to complete the classification task with maximum accuracy—62.5% when trained on all three categories and 90.5% when the Classical and Romantic labels were merged. These results indicate that the model was most effective at distinguishing Baroque music, suggesting clearer stylistic separation. These findings demonstrate the feasibility of using sentence-level embeddings for symbolic music classification, offering potential applications in musicological analysis, genre tagging for recommendation systems, and quantitative exploration of musical style beyond human perception.