Audio should be easier. You could for example only transcribe voices that match a voice that has given permission in that voice. You’d still need AI to analyze whether it was AI generated but I suspect doing that on voice is easier than text.

→ View original post on X — @yoheinakajima