Got a notice about Kickstarter affiliated with OpenSubtitles for an on-demand AI subtitle generator and player that claims it supports Plex and Kodi and Jellyfin.
Anyone else heard about this yet and how it would work within Kodi?
Sounds like they will somehow handle this via an add-on, but I’m curious to know how much of this is real versus another make-a-quick-buck AI-hyped play?
@sam_nazarko has OSMC been contacted by this group?
https://www.kickstarter.com/projects/techspecs/ray-watch-anything-understand-everything
I know nothing about that particular project but opensubtitles web sites will already AI translate subs to other languages. You can download a free program called SubtitleEdit that has lots of options for doing text to speech conversions. Seems to me that someone would need to be in the situation where they were watching a lot of content that happened to have no subtitles available online for this to make any sense bothering with let along investing in.
Regarding the AI aspect generally, opensubtitles themselves will discount or refuse subs for being machine translated. My personal experience making subs for a Japanese film with only Japanese subs available (I only speak English) is that machine translating has a long way to go. When I started with the normal translation apps what came out had many lines that clearly were not correct. When I made another go using tools that gave alternate translations to choose from I got closer but their were some parts I couldn’t quite figure out the full context of the conversation. I then ran it through a couple AI models to translate it and it didn’t do any better than what I had already gotten to, but also inserted some text that was clearly wrong as it incorrectly made assumptions on who said what which changed the context of some conversations. It wasn’t until I spent hours with the AI going back and forth describing scenes and asking and answering questions until I got an accurate translation out of it. I’m rather sceptical currently anyone has an AI model that comes anywhere close to what a human can do.
1 Like
Haven’t been contacted by them.
It’s doable – the problem with lower powered systems: either you would be using bandwidth and sending an entire bitstream which is tricky, particularly when handling seeking or you do it on device.
I can see from their Kickstarter that they intend to use the models on device from their recommended HW requirements.