@meowski >all but the simplest ones, could be seen as proprietary blobs that are difficult to reverse engineer
Yes, that is another problem.

How can you be sure the software serves you if the software cannot be understood even collectively?


Training speech models is unnecessarily complex - what you really need is a voice actor with a clear voice and to get them to pronounce all the English sounds and a few hundred or thousand of the most common words and tag every single one and then you'll be able to programmatically output whatever sentences you want with only uncommon words sounding odd - alas nothing like that has been released under a free license.