The only way I can think of this being justified is if it's run and controlled by the user themself so they can control what they hear. Not to control how others speak.
Moderation is trying to enforce some standards for the whole environment. Even if you got the tools to implement filtering for yourself, you're still sitting there with people who are happy with the abusive environment and will get through one way or another. And you get them everywhere around you. It's not something that each player can deal with separately unless they opt into private groups - but that's not the environment Roblox is encouraging.
E.g. while it's not perfect, Valorant's audio chats are a lot less toxic than many comparable FPS games via such a system. Of course that still won't protect you from one-off hearing slurs.
So that + elective client side filtering (if you really never want to experience profanity), seems like the best of both worlds.
manual reports + reviews requires recording the feed for later review; real-time filtering does not. A number of other commenters have mentioned that being recorded bothers them; those folk should be arguing for the real-time AI moderator.
This absolutely will be used for surveillance. It's literal express purpose is surveillance and automated behavior modification.
Basically, if anyone wanted to do surveillance like this, they were always doing it with public or secret implementations. This won't massively help anyone.
Imagine a place on the internet where your speech is moderated, like a place you’re not forced to use and has rules about what you can and can’t post. Speech that goes against the rules could be hidden or removed and the users of the service train themselves to speak in a certain way when they are using it.
Perish the thought, that sounds like a dang dystopian Netflix show
But yes, please keep defending the development of mass surveillance technology.