Not my field of expertise, but seeing that generating the voice is one of the biggest costs, have you think about generating it yourself with something like https://github.com/mozilla/TTS running the process in commodity hardware?