The 80% solution: Pair with a professional narrator who has consented to have their voice modeled by this (see the note at the bottom about what they held back from open sourcing). This generates a beta, and then you can pay the human narrator to rework specific sections you’re unhappy with.