tone generated by analog circuits is notoriously difficult to reproduce digitally. the conceptual behavior is often easy to model, but hardware deviates from theory in ways that are technically subtle but audibly apparent.

it's deterministic, but the parameters may be unknown and approximate values must often be discovered by iterative guess-and-check. researching and manually modeling an approximation can be incredibly tedious and still fall short. this is exactly the kind of application that machine learning excels at.

Not a silly question. One of the things I always miss from these studies is comparison with more classical inference models. In particular, back in the days we had to train against outputs for PCA models (linear and non-linear) in order to get a feasible production version with the best reduced parameters and justify equivalent performance. Today we have so much computing power that nobody cares.

Audio is a particular good application of this. For example, the old ADPCM algorithms have evolved naturally into their ML counterparts. Some have even less parameters and thus are more computationally efficient because of the advantages of flexible feedback of the training or production models (e.g. RNNs).