A lot of these are tricky files , for example the Agemix files, these are almost always compillation videos which edit out all of the bits with dialog and just get to the ,, OH you have to forgive me for this awful pun, but they just get to the climax. So you could have a 4 Hour, 8 Hour, etc. video with almost no dialog so even the most finely tuned Whisper is still going to end up with those damned hallucinations. I'm always tweaking whisper, I know that you know the story, but without knowing Japanese and with over 20Thousand files to deal with I have to just take the best that I can get. Presently I'm mostly happy with the results that I am getting. One little issue is that occasionally whisper/VAD will completely ignore dialog that is obvious, clearly spoken and without any surface noise. I have tweaked and tweaked and tweaked my files to try to resolve this but the trade-off is more hallucinations, I hate hallucinations so my current configs are set to agressively remove hallucinations and if there are bits of dialog missed, shrug... Oh, One other thing, it would be interesting to reevaluate these files after my re-scan of everything, If you're up to it, see what you find after I've run everything through again, I expect to be done with 'A' maybe Tomorrow if all goes well, I have a ton of other background work todo, renaming all my files, I'm extracting all of the useless text after the filename, One problem with that is I am blanket-removing all text past 11 Characters, 11 because there are files named longer than others so I don't want to remove some of the code but the end result is I'm getting a lot of files named, ASW-234 de , The de is a result of that 11 character issue, most of the new files will have this but at least all of that idiotic long text will be gone. IN any case, I've babbled too long, Thanks for the excellent feedback Mei.
I use renamer. I was unaware taht Javluv could rename correctly, My problem is there are codes with 7 characters all the way up to over 20 characters, like the Caribbean.com files or the fc2-ppv-xxxxxxxxxxx but most are xxx-xxx or xxx-xxxxx or xx-xxx. there are all kinds of variations in the names so it is , from my knowledge base, only possible to generalize if you're removing characters from a spacific point. Are you saying that JAVluv can batch rename like that? I don't want to rename each file individually, that would take forever. I want to rename by folder, for istance, earlier tonight I renamed all of my "M" files, that's over 8000 files, some wiht 6 characters, some with 7, some wiht 8, so the way, again I may be ignorant of the possibilities, but to rename everything and not lose characters from files with longer names is to set my renamer to remove all characters beyond 11. Net result, some have a few extra characters but if you know a way to batch rename and have them all correct, fantastic. IN the mean time I will try and figure out what you're referring to with Javluv. Thanks Mei
Confirmed: I was able to download the subs (google drive zipped).
I have a program that scan the folder, find possible movie information on internet and export to an HTML file with title and image. I will try with your folder in a few days. Currently I'm busy with some improvement.
Hiya Theydon... that looks like an interesting program. I will look at your html file some more. I don't think that it would solve my naming issue but this looks like it may have other uses. You may want to point this to Tmpguy, who developed Javluv (do you know about JAVLUV?, if not, you need to get it) Search for Tmpguy on AKIBA-online and at the bottom of his posts is a link to javluv. Definitely check that out. That said, the movie information is very interesting. Thanks for sending this. Cheers.
There are a Million little complications with Whisper. My .bat and .py files scan a video first with Vad-True/Large-v3 and all kinds of tweaks to VAD filters but at some point the intent to catch whispered dialog gives you Hallucination. Carrying on, if VAD-True yields a .srt less that 10K, then it doubles back and re-scans the file with VAD-False and Large-V3. It is working reasonably well but there's always little issues that pop up here and there. I don't think that you can get perfect unless you manually edit and you have to know Japanese to do that. I hope that my files will be 'good enough' for most people but they will never be perfect.. Cheers/
mdx_kim2 could help with that, it's a better voice extraction method than what faster-whisper uses (which is HT demucs I think ?). I use whisperXXL
