Building a Dataset for Music Analysis and Conditional Generation

This is my undergraduate thesis project. Current deep learning music models are able to create expressive music but offer little control over its output, mainly due to their inability to understand musical structure. To address this, we built a dataset of piano music containing additional tracks of annotated structural information such as melody, beat positions, phrase locations, etc. This allows us to explicitly condition on musical structural features, opening up the possibility of improvements in controllable music generation and deep learning-based melody prediction.