Transfer Learning for Underrepresented Music Generation

Abstract

The increasing popularity of Deep Neural Networks (DNN) has led to their application to many domains, including Music Generation. However, these large DNN-based models are heavily dependent on their training dataset, which means they perform poorly on musical genres that are out-of-distribution (OOD) for that dataset. This heavily limits these systems' practical use and essentially requires the model to be retrained on a large dataset containing a musical genre in order to recreate it. In many domains, transfer learning has been effective at adapting an existing model to a new target dataset of a much smaller size by training for a much shorter period. However, such an approach remains underexplored in the domain of music generation. To investigate the viability of this approach, we explored different genres that might represent OOD genres for a DNN-based music generator. Consequently, we identified Iranian folk music as an example of such a genre of music. This was in line with the fact that this genre has a melodic structure different from music based on Western music theory principles. We then proceeded to collect a dataset of Iranian folk music and utilize it to explore different methods of transfer learning to improve the performance of MusicVAE, a large generative music model with a DNN architecture. We identify a transfer learning approach that allows us to efficiently adapt MusicVAE to the Iranian folk music dataset, which indicates a potential for the future generation of underrepresented music genres.