Training a model on sound data from readily available public social network posts and targeting their followers (which on say fb would include family and is full of "olds") isn't a very far fetched use-case for AI. I've created audio models used as audiobook narrators where you can trivially make a "frantic/panicked" voice clip saying "help it's [grandson], I'm in jail and need bail. Send money to [scammer]"
If it's not happening yet, it will...