Yes! And not just notes or chords or time signatures, but real analysis of how the piece works.
I believe parts of this model could be adapted to accomplish this. The "mulan" network generates embeddings from music and text so that they "match" each other, i.e. audio segments and text captions that go well together would be mapped to the same embedding. At run-time, they use embeddings from the text to generate the music. So conceivably the opposite could be done, using the "mulan" embeddings from audio only to generate text, by replacing the "soundstream" decoder with a text LLM. That said, there's probably simpler methods that could generate decent results, and I'd be surprised if there isn't already some out there. It's worth noting that along with this work they released a dataset of music with captions that could also help training new models for this task.