> This has some truly weird implications, e.g. once language model translations switch from subtitles to dubbing it will effectively stop shifts in pronunciation, because the models won't be re-trained with an ear to the street.

This has already happened to a great degree thanks first to radio and the television, as well as the early 20th century movement to “received pronounciation” in many countries. E.g. TV finally killed off thee and thou in the 1960s

It often feels like there are regional accent variations but they are usually quite minimal these days thanks to spread of technology.

The in Europe, explicit suppression or uniformation of language began AFAIK with Louis XIII and his deliberate formation of “France” (as opposed to just a collection of regions controlled by one person). In China I believe the same thing was instigated by the (by coincidence contemporary!) Qing dynasty, but it might have been a lot earlier. In any case it really zoomed throughout the world in the 20th century when communication technologies and practices were adopted by the emerging nationalist movements.

So machine translation will simply continue a longstanding process.