Yes, I'm sure whatever TTS model Replica uses was only trained using voice data provided by opted-in, informed consenting voice actors
This is the exact same AI bullshit every other AI company is doing, using scraped data from millions or billions of users en masse (or re-using models trained thusly) but then for the VERY last step of fine-tuning the model to a particular voice they slap on the absolute thinnest layer of ethics, saying they'll only host fine-tuned models trained on VAs who've signed with them.
Fuck this. Fuck Replica. And Fuck SAG-AFTRA for going along with and legitimising it.
This is like taking StableDiffusion, and then training a LORA or an embedding with art you have specifically licensed for AI training use and calling the end-result and images generated with it "ethical".