I'm not a data science / AI person at all, but AFAIK while initial model training is enormously processor intensive, customising a trained model is not. This ha

I'm not a data science / AI person at all, but AFAIK while initial model training is enormously processor intensive, customising a trained model is not. This has been my experience playing with custom training stable diffusion on very low memory home GPU. I'd assume it's true for voice generators like Tortoise [1] also. Moreover, while Eleven Labs isn't in my opinion as good, they let you custom train voices for an incredibly low cost and with a tiny amount of sample data [3]. For perfect audio spoofing, lots of audio would be needed. Bear in mind there are literally millions of podcasts available [2], and billions of youtube videos. Should be trivial to grab biographic data and voice samples from a subset of them. [1] https://github.com/neonbjb/tortoise-tts https://github.com/neonbjb/tortoise-tts [2] https://whatsthebigdata.com/podcasts-statistics/ https://whatsthebigdata.com/podcasts-statistics/ [3] https://elevenlabs.io/ https://elevenlabs.io/