I'm not sure how useful any of this is anyway though. I took another look at Amazon Transcribe a couple weeks ago and reached the conclusion that its transcriptions are still pretty bad compared to a human doing it. You get the general gist so long as it's a clean recording but there are lots of errors.

Standalone software is going to be even worse.

Imperfect captioning is better than nothing in some cases. But who wants to read a book--much less use it as a learning tool--when there are errors every other sentence?

ADDED: It really doesn't seem like much of a threat to publishers in its current form. Probably lots of errors and the text is only delivered a few lines at a time as the book is read aloud. However, I can see publishers looking down the road and imagining much improved ML transcription and larger blocks of text as something that's legally indistinguishable.

Really the only actual answer to that question is because a lawyer convinced a judge to say that they can't. If they can get a better lawyer and/or a different judge then maybe all of sudden they can.