Have you considered using 'learning from human preferences' as the loss function in addition to the Transformers? That was another OpenAI project, and it seems tailor-made for music generation: what is more 'I know it when I hear it' than music quality?