I wonder if author could follow something similar that is used to capture guitar amps to NAM models (Neural Amp Modeller), https://www.tone3000.com/capture

I've reconstructed everything I'd wanted so far, but I'll keep an eye on this. This also has an option to learn from just a recording of its test sweep (like the technique I used). I suppose a learned model could be useful for nonlinear modes, allowing a more complex function when my current approach fails (which is quantifiable with the existing tests).

Also my experience

Hmm word2vec doesn’t seem to belong to any on the four. Dean was a last coauthor of the paper.

You're looking for this paper:

https://arxiv.org/abs/1405.4053