[smt88]

First: this isn't a thought exercise. People have been doing it for years[1][2] and are trying to commercialize it. The results have become frighteningly convincing, at least for some people. You can convincingly clone both voices and writing, and the combination is even more convincing.

The training set would be exactly what you mentioned: chats between you and your family, friends, and coworkers. Professional writing might be useful for another purpose, to replicate your writing style (as we've seen done to famous authors).

Most people send hundreds of personal messages a week, which (over time) is more than enough to train a model.

1. https://www.technologyreview.com/2022/10/18/1061320/digital-...

2. https://economictimes.indiatimes.com/magazines/panache/miss-...

[csw-001]

This is really interesting - Thanks for sharing! It looks like these services are taking an "interview" approach, so they are building their own dataset. That makes a ton of sense, and really avoids a lot of the concern of cloning people without their consent (it would be hard to talk them into hours of interviews without their knowledge).

[dmarchand90]

All of this reminds me of the the old fake infomercial "live forever as you are now" https://youtu.be/xg29TuWo0Yo?si=yvFa-vFIEqD-Swx4