Yea, it's a good point. We can logically distinguish between the basic audio problem (which might be really hard) from the automation problem.
On the other hand, suppose we somehow got good training data by getting a bunch of audio samples at the same number of words per minute that were graded by human listeners as easy or hard to understand. Then in principle something like a neural net might figure out what audio features were responsible for intelligibility and then adjust the non-intelligible audio in that direction (a la using convolutional neural nets to make pictures appear in the style of a famous painter without changing the content). This would be done automatically without any humans actually understanding the solution.
Sure, you could try any number of things to produce a solution. But even if you try an approach where you don't know at first what might work, you should likely still put effort into figuring out what features made it work, so that you can improve it further and maintain stability.
On the other hand, suppose we somehow got good training data by getting a bunch of audio samples at the same number of words per minute that were graded by human listeners as easy or hard to understand. Then in principle something like a neural net might figure out what audio features were responsible for intelligibility and then adjust the non-intelligible audio in that direction (a la using convolutional neural nets to make pictures appear in the style of a famous painter without changing the content). This would be done automatically without any humans actually understanding the solution.