Contrary to some posts here, I think this approach shows promise though it would clearly benefit from further refinement.

The criticisms here seem to break down into three groups:

1. "I don't mind reading subtitles and that will always be better."

2. "Any real film purist will prefer watching the original."

3. "The implementation is kind of uncanny at some points."

My thoughts:

#1: Sure, I actually agree but I also know that there are a lot of people out there that, for whatever reason, find reading subtitles breaks their immersion. Even though that's not me, dismissing that audience's experience isn't very inclusive.

#2: Yep, As a film maker I totally get that but us purists and aficionados aren't that big of a market. So yeah, Criterion should never use any tech like this for definitive remaster releases as it doesn't preserve artistic fidelity. But just because they should keep this away from Citizen Kane doesn't mean there aren't suitable classes of content. In the era of high-fidelity streaming media, additional versions don't necessarily take distribution availability from the original (which should always be available).

#3: Though I'm not an AI engineer I've done some work with AI engineers on artistic content and my guess is the current gaps we see might be sufficiently surmountable with refinement. It may never be fully automatic but redubbing is human intensive anyway so an interactive UI that allows an ADR-experienced editor to tune problematic output may help a lot.

I think the team should keep going. Once it's good enough, there is a lot of content that might reach new audiences through this tech.