Controlling the Emotion of Music using Generative Deep Learning

This research project aims to develop novel, knowledge-based deep learning techniques for emotional music generation. The primary objectives are: (1) To explore effective methods for converting 'vanilla' melodies into ones that express a specific emotional state, and (2) To identify which technique among statistical analysis, training neural networks, music information retrieval (MIR), or music theory offers the best emotion control over generated melodies.

The approach will involve resampling a melody, generated by an existing model, note-by-note, to reflect the chosen emotion. The effectiveness of each technique will be empirically evaluated using small datasets of symbolic music. A prototype will be created to allow users to either generate a new melody with a specified emotion or convert an existing melody to the chosen emotional state.

This project holds novelty in applying knowledge-enhanced deep learning methods to music composition for emotion representation. Its contributions will expand the knowledge in knowledge-based deep learning techniques and underscore the intersection of engineering, physical sciences, and musicology.