News for hackers. Stuff that matters.
Re:Seamless: Meta's New Speech Models by mrob on Friday December 01, 2023 @11:08AM (#38488312)
Differences in verb-subject-object word order will always add latency. If you want to translate from German, with the verb at the end, to Welsh, where the verb goes at the start, you'll have to wait for the complete sentence before you can begin.
Re:Seamless: Meta's New Speech Models by d3m0t3p on Friday December 01, 2023 @11:31AM (#38488614)
Not necessarily true, for the first few sentences you won’t be able to do it. But afterwards, once the context is established you don’t really need to wait for the verb, you can predict it. For example if you are speaking about cleaning the house and you detail that you have cleaned the kitchen the stove and so on, you can predict the verb with only the start of the sentence. I don’t have any source to back this up, but it sounds plausible
Re:Seamless: Meta's New Speech Models by gberger on Friday December 01, 2023 @11:53AM (#38488933)
What if the predicted verb was incorrect, but the model has already translated the incorrect prediction? How does it tell you about a mistake?
Re:Seamless: Meta's New Speech Models by mrandish on Friday December 01, 2023 @12:44PM (#38489629)
A good approach might be to start with how top notch, ultra-experienced human translators handle corrections for real-time scenarios, for example, the expert translators that do the ear monitors at the United Nations. I've worked with a few such real-time translators when preparing keynote speeches and they seem to have rigorous processes that appeared quite deep. Probably a ton of domain expertise to be captured there.

That said, I suspect that real-time language translation is always going to be somewhat imperfect due to its nature. Non-real-time translation of literature is still a subjective art form even at the very high-end of human expertise.

Re:Seamless: Meta's New Speech Models by shkkmo on Friday December 01, 2023 @12:35PM (#38489518)
Once you start predicting what someone is going to say you are no longer translating their speech
Re:Seamless: Meta's New Speech Models by Teever on Friday December 01, 2023 @04:13PM (#38492483)
Yeah but then you're just introducing branch mispredictions which will cause latency and potential confusion down the line.

It's all a trade off.

Either way it's extremely exciting that we get to even discuss this stuff as real possiblities.

Re:Seamless: Meta's New Speech Models by tralarpa on Friday December 01, 2023 @11:21AM (#38488467)
It's very impressive what simultaneous interpreters can do. They don't wait for the end of the sentence.
Re:Seamless: Meta's New Speech Models by MrsPeaches on Friday December 01, 2023 @11:43AM (#38488793)
Even they struggle with jokes though.

This may be apocryphal but I’ve heard that in formal settings (e.g. UN) they won’t translate it and will instead give instruction on when to laugh.

Re:Seamless: Meta's New Speech Models by numpad0 on Friday December 01, 2023 @11:35AM (#38488690)
Yeah they backtrack on branch prediction failures.
Re:Seamless: Meta's New Speech Models by dylan604 on Friday December 01, 2023 @11:39AM (#38488747)
What kind of heartbleed that must introduce.
Re:Seamless: Meta's New Speech Models by Vecr on Friday December 01, 2023 @12:37PM (#38489539)
You mean meltdown/Spectre?
Re:Seamless: Meta's New Speech Models by dylan604 on Friday December 01, 2023 @12:55PM (#38489841)
probably, but you got the gist anyways