The trick is not to use NNs for DSP but to discover the parameters for the DSP. In other words you hardcode the signal flow architecture using a common technique like an FDN but then train a NN to find "good" sounding parameters, like comparing to a convolution reverb or recordings.

The thing about reverb is they require a lot of state and nonlinearity is undesirable.

For reverb I don't see much practical use, mainly because you can capture a pretty-much-perfect recreation of a real space with an impulse response. No need for thousands or millions of rounds of training a network. For unrealistic reverbs, you have the problem that to get training data you'd have to invent several unrealistic reverb effects to apply to sounds. And once you've made those effects, there's not really any reason to neural-netify them.

For NeuralDSP it's a bit different because they use NN's to simulate a guitar amp circuit which is a nonlinear system and so there's no simple way to "capture" the effect the way that you can for reverb sims or speaker sims. And while you can make a very accurate model using something like SPICE, that won't run in realtime. With traditional amp modeling you basically take the SPICE version and try to optimize and cheat as much as you can so it can run in realtime, at the cost of accuracy.

So that's what NeuralDSP's goal is, a system that approximates the amplifier but can also be computed in real-time, except done using a trained NN instead of a human-optimized variant of the SPICE circuit.

They have a couple whitepapers on their website, though none of them go deep enough to really give away their secret sauce. But basically according to them, making a NN model of an amplifier at a fixed setting is fairly simple. Where they had to get novel with it is adjustable settings/parameters. E.g. turning the drive up, or turning the treble down. Just capturing a few hundred or thousand models based on adjusting parameters and cross-fading between them doesn't sound realistic. So they had to come up with a larger model architecture that can "learn" those parameter changes.

https://arxiv.org/pdf/2403.08559

Not sure about Neural DSP or reverbs in general, but real-time neural network based DSP seems very possible. The open source Neural amp modeler[1] would be a good place to start diving in.

[1] https://www.neuralampmodeler.com/the-code