I don't know if you have kids/grandkids/nieces or nephews (or plan to have those) but it might be nice to record your wife reading some books out loud.
Not only will you have your own personal "audio books" of Harry Potter/The Hobbit/Chronicles of Narnia/Oi Frog/Alice in Wonderland/Roald Dahls etc etc for any kids/grandkids/relatives etc that will hopefully be something treasured in its own right, but you'll also have a large corpus of training data from well-known texts that you can retrain over and over as the tech improves in the future. Might be worth chucking in some other well-known texts to avoid over-fitting on a "kids' story voice" - maybe something plain like inauguration speeches/declaration of independence/magna carta/etc.
Obviously I'd focus on gathering raw material now, and focus on the reconstruction later when you've all recovered mentally and physically to whatever happens. The more data the better when it comes to this sort of thing. There might not be something "simple" right now (e.g. you could probably implement the WaveNet or similar paper yourself today, and training it up on some GPUs in your spare room etc, but in a few years there might be a nice WYSIWYG/SaaS thing for it), but with the recordings safely stored you'll obviously be able to use it in the future.
I like this idea but the specific examples you give would almost certainly be a terrible idea. A voice trained on Tolkien or old American legalese like the Magna Carta would train a model with a lot of thee, thus, therefore and though art and undertrain it with modern English. His wife would sound like the second coming of Jesus or Shakespeare and less like a normal human being.
From what I understand, it is not the words themselves (thee etc) but the sounds that make the words - so the "th" and the "ee" are still legit sounds in modern English words. The network would just be synthesising the words you tell it to - it won't be picking the words for you.
Not only will you have your own personal "audio books" of Harry Potter/The Hobbit/Chronicles of Narnia/Oi Frog/Alice in Wonderland/Roald Dahls etc etc for any kids/grandkids/relatives etc that will hopefully be something treasured in its own right, but you'll also have a large corpus of training data from well-known texts that you can retrain over and over as the tech improves in the future. Might be worth chucking in some other well-known texts to avoid over-fitting on a "kids' story voice" - maybe something plain like inauguration speeches/declaration of independence/magna carta/etc.
Obviously I'd focus on gathering raw material now, and focus on the reconstruction later when you've all recovered mentally and physically to whatever happens. The more data the better when it comes to this sort of thing. There might not be something "simple" right now (e.g. you could probably implement the WaveNet or similar paper yourself today, and training it up on some GPUs in your spare room etc, but in a few years there might be a nice WYSIWYG/SaaS thing for it), but with the recordings safely stored you'll obviously be able to use it in the future.
Best of luck to you both.