It is probably good that OS community are exploring this however I'm not sure the technology is ready (or will ever be maybe) and it potentially undermines the labour intensive activity of producing high quality subtitling for accessibility.
I use them quite a lot and I've noticed they really struggle on key things like regional/national dialects, subject specific words and situations where context would allow improvement (e.g. a word invented solely in the universe of the media). So it's probably managing 95% accuracy which is that danger zone where its good enough that no one checks it but bad enough that it can be really confusing if you are reliant on then. If we care about accessibility we need to care about it being high quality.
They're helpful to my deaf ears, even when they're wrong (50% of the words) they do give me a solid idea of what is being said together with what the audio sounds like.
With it, I get almost everything correct. Without it, I understand near to nothing.
This only goes for English spoken by Americans and sometimes London Britons, sadly, nothing else get detected nearly as good enough, so I can't enjoy YouTube in my native language (Dutch), but being able to consume English YouTube already helps a lot!
I've been messing with more recent open-source AI Subtitling models via Subtitle Editor which has a nice GUI for it. Quality is much better these days, at least for English. It still makes mistakes, but the mistakes are on the level of "I misheard what they said and had little context for the conversation" or "the speaker has an accent which makes it hard to understand what they're saying" mistakes, which is way better than most YouTube Auto Transriptions I've seen.
This is not by default bad thing, if it is something you only use when you decide to do so, when you don't have other subtitles available tbh.
I hate AI slop too but people just go to monkey brain rage mode when they read AI and stop processing any further information.
I'd still always prefer human translated subtitles if possible.
However, right now I'm looking into translating entire book via LLM cause it would be only way to read that book, as it is not published in any language I speak. I speak English well enough, so I don't really need subtitles, just like to have them on so I won't miss anything.
For English language movies, I'd probably just watch them without subtitles if those were AI, as I don't really need them, more like nice to have in case I miss something.
For languages I don't understand, it might be good, although I wager it will be quite bad for less common languages.
Personally, I would be happy even if it didn't translate it but were able to give some half decent transcription of, at least, English voice into English text. I prefer having subtitles, even when I speak the language, because it helps in noisy environments and/or when the characters mumble / have weird accents.
However, even that would likely be difficult with a lightweight model. Even big companies like Google often struggle with their autogenerated subtitles. When there's some very context-specific terminology, or uncommon names, it fumbles. And adding translation to an already incorrect transcript multiplies the nonsense, even if the translation were technically correct.
There are a number of open weight open source models out there with all their data sourced from the public domain. Look up BLOOM and Falcon. There are others.
JetBrains' AI code suggestions were only trained on code where authors gave explicit permission for it, but that's the only one I know from the top of my head.
Most chat-oriented LLMs (ChatGPT, Claude, Gemini...) were almost certainly trained using corporate piracy.
It won't be better than human translated ones but begter than no subtitles. I don't think even humans can make subtitles correctly without knowing context
Honestly, if it can generate subtitle files it'll be a huge benefit to people creating subtitles. It's way easier to start with bad subs and fix them than it is to write from scratch.
Im curious What makes what VLC is doing qualify as artificial intelligence instead of just an automated transcription plugin?
Automated transcription software has been around for decades, I totally understand getting in on the ai hype train but i guess I'm confused as to if software from years past like "dragon naturally speaking" or Shazam are also LLMs that predate openAI or is how those services worked to identify things different from how modern llms work?
Llms are a very specific Gennerative AI subset. Not everything AI is LLM, especially stuff like Shazam is pretty traditional AI. It's been around for a while already, and studied for even longer (even back in the 1960s we were already starting to have a field of study in this domain)
VLC always had a ton of applications, network device playback, TV, streaming server, files, physical media, music player, effects, recording, AV format conversion, subtitles, plugins and so on.
"Do one thing well" is what gives you software like sendmail, which requires several other programs to be actually useful, all of which have to be configured separately to work together, with wildly different syntax.
Cause we can no longer sit back and allow AI infiltration, AI indoctrination, AI subversion and the international AI conspiracy to sap and impurify all of our precious bodily fluids.