i love those before-and-after interviews. it was probably all just for the camera, but it worked on me too lol. also, storylines like “your gf won’t swallow? i will do it then” like in MIAA-735 at 17m12s. this part is hard for AI to get right, so i usually use it as a benchmark to get an idea of how good a model or my settings are. it probably means nothing, because doing well here doesn’t guarantee the same results for the rest of the video or for other videos.Yeah, I totally agree, What is "good enough" for me? That is the Million dollar question. And as I said in my over-long description, the vast bulk of JAV 'dialog' is just the same thing over and over and over. For me, the only time dialog matters is during interviews or to a large degree the pornless scenes. These scenes I am often curious about the comments. My 'fetish' I suppose you could call it, is post action interviews. I am interested in what the the girls have to say after they've just been bukkaked or swallowed a ton of cum. If anyone has some suggestions On that score I would be happy to have them, But back to the point, YOu are completely right Krack, Just about all of most JAV titles don't really need comprehensive subs. There is probably about 1% of all of the titles where I even give a fuck what is being said, otherwise I already know what is being said.. However, my objective doing these was not so much to please myself as it was a project to contribute to the community.
from my limited testing, i feel like offline methods are already good enough for speech detection and transcription. there aren’t that many things that can go wrong there. the real limitation seems to be translation with quantized offline models. so uploading jp subs and letting cloud AI help with the translation/context part might give better results. just a suggestion for anyone who hasn’t tried this setup yet.

