A Predictive Neural Network for Tonal Melodic Completion

Philip Baczewski, Independent Scholar

This study shows the utility of a predictive neural network to help understand the influence of structural and theoretical elements on musical style acquisition. The task modeled was the note-to-note prediction of pitches within simple melodies of a common practice tonal style. By using the neural network as an analog for human musical style acquisition, it is possible to explore the influence of structural or theoretical elements on the level of predictive success. The model was able to predict melodic notes with a better than random accuracy at over 55% for some individual melodies. The model also showed an ability to follow melodic contour. An additional prediction task was performed using selected melodies from the works of Beethoven. Results showed a predictive accuracy of over 50% on individual melodic examples indicating predictive power across common practice style periods. Neural network technology shows promise in helping validate music theoretical concepts by serving as an analog to culturally mediated learning of musical style.