The protocol for sensory input has likewise been prosthesised by Paul Nach-y-Rita’s tongue grid array for vision.
https://www.researchgate.net/scientific-contributions/Paul-B...
What’s left is all the logic between input and output.
We gain a pathway for investigation backwards. What neural substrate is active in generating the speech data? Work backwards from there.
We already have a useful pathway for investigation: data from speech/sign of a speaker of a particular language. In fact this already yields a kind of overabundance of data. The tricky part is finding the right kind of data via careful experimentation and organising it, by building explanatory theories of language that meet the conditions of evolvability and learnability.
(Which is not to say that cognitive sciences can't shed any light on this, but studying the brain independently of linguistic experimentation is a dead end IMO. Brain imaging studies coupled with linguistic studies have yielded interesting results - see Andrea Moro's work on "impossible languages").
Actually we have some idea and one theory is that they are not as distinct systems as we’d like to think, at least they work bidirectionally and reciprocally: language structures back the thoughts themselves, so there is a way that externalization actually feeds “inwards”. See distributed/embodied cognition.
As for the popularity of linguistic determinism, who knows. I wouldn't trust a linguist who took the strong Sapir-Wharf hypothesis seriously.
Being a universal function approximator doesn't magically solve every problem.
More importantly, the fact that any function can be represented as a NN does NOT mean that any function can be learned through known NN training mechanisms, not even in principle (i.e. given finitely arbitrarily many examples and unbounded but finite time).
And of course, there is always still the vague possibility (which I don't personally subscribe to) that the actual function is not Turing computable, which would mean it certainly can't be approximated by an NN.
There has never been a dearth of universal function approximators, polynomials can do it, splines can do it, sine/cosines can do it. Being a universal approximator is hardly unique or special.
There is absolutely something special about DNNs but being universal approximators is not one of them.
Being able to learn a function from data is very different from being able to represent that function.