Composer classification: neural networks and harmonic analyses

This project offers a new approach to the task of computationally predicting which composer wrote a piece of music. For this task, two types of neural networks are tested: an LSTM network and a GRU network. Previous studies approached this task using either digital sheet music or audio recordings as input. This project introduces the use of harmonic analyses (chord progressions). My hypothesis was that this information as input would be sufficient for this task. Three composers (Bach, Beethoven & Monteverdi) are represented by 45 pieces each. This is split into two collections: a large collection (80%) to train on, and one that remains ‘hidden’ during training to test the model afterwards. All of these new pieces are classified correctly by my model, outperforming previous techniques.