Post your JAV subtitle files here - JAV Subtitle Repository (JSP)★NOT A SUB REQUEST THREAD★
- Thread starter Eastboyza
- Start date
-
Akiba-Online is sponsored by FileJoker.
FileJoker is a required filehost for all new posts and content replies in the Direct Downloads subforums.
Failure to include FileJoker links for Direct Download posts will result in deletion of your posts or worse.
For more information see this thread.
If you do whisper --help it'll show you all the options and how to use them(kinda).
This is part of what it says:
Doesn't say how it can use multiple but I'd assume something like whisper --temperature (0.2, 0.4, 0.5) from your python example. Maybe " instead of ( and ).
I haven't messed with extra options at all so no idea how they work.
This is part of what it says:
Code:
--temperature TEMPERATURE temperature to use for sampling (default: 0)
--best_of BEST_OF number of candidates when sampling with non-zero temperature (default: 5)Doesn't say how it can use multiple but I'd assume something like whisper --temperature (0.2, 0.4, 0.5) from your python example. Maybe " instead of ( and ).
I haven't messed with extra options at all so no idea how they work.
I used DeepL pro (API calls) for IPX-998 translation --I too was surprised how good the translation became. I did not expect such a difference between pro (API calls) and the web version. In terms of manual post-edits, I needed to do roughly 10 edits or so.
I looked at the help, and tried all combintaions of "" (), and []. Searching for help with Whisper online is just extremely difficult because it is new, the name sucks, and because virtually nobody seems to be using the command line. https://blog.deepgram.com/exploring-whisper/ has examples for Python with:
beam_size=5
best_of=5
temperature=(0.0, 0.2, 0.4, 0.6, 0.8, 1.0)
so I guess that makes temperature a Tuple of Integers? Seems really sloppy on how that interacts with command line. Since the command line --help doesn't even give the suggestion of multiple values, it just might not be possible.
Have you tried with nothing around? So like --temperature 0.2, 0.4, 0.5 or even --temperature 0.2 0.4 0.5
cmd> whisper --model tiny --language ja --task translate --temperature 0.4 --temperature_increment 0.1 .\inputfile.wav
DVDES-626 Sex Education 4 For I Want To Tell The Son Of Incest Planning Beloved Issues In The Ultimate To Be ... Pregnancy
I used Whisper to produce this subtitle file for DVDES-625. A word of caution: I downloaded the first part and then down loaded the second part from a different site so the timing might require adjustment. As always however, I still had to clean it up a bit and re-interpreted some of the meaningless/ "lewd-less" dialog. Again, I don't understand Japanese or Chinese so my re-interpretations might not be totally accurate but I try to match what is happening in the scene. Anyway, enjoy and let me know what you think.
DASS-091 My Son Is A Libido Monster --Hana Himesaki
This is a file I got from subtitlecat. It was already very good, so I just polished it up a bit mostly by fixing pronouns and dealing with all the "so comfortable" comments. I also chose to substitute the more natural sounding "sex fiend" for "libido monster." This is one of the subgenre of JAVs in which a mother doesn't know what to do with her continually-masturbating son who is so horny he is beginning to get tired of his blow-up doll and starting to eye mom lasciviously. In this particular case mom is in luck because a young woman living in her building is also a sex fiend and offers to help out. For a darker and more twisted take on the same theme, check out the recently released HUNBL-126. As far as I know, no subtitles exist yet for that one, so if anyone is feeling ambitious it might make a good project.
Are you using a legit adobe premiere pro or just a cracked one?
I am planning to try it myself but I don't want to pay the expensive subscription of adobe products.
PiratedAre you using a legit adobe premiere pro or just a cracked one?
I am planning to try it myself but I don't want to pay the expensive subscription of adobe products.
Has anybody tested how whisper translation compares to deepl (free version)? Up to know I found whisper doing a very good job but since I can't save a japanese version for deepl and let whisper translate it for he same file, it's hard to tell.
change the option of translation_mode : No translation
I know how it workes, the question is if you get better results using whisper or deepl.
This will not work as Whisper generates different subtiltes every time you run it, so there is no one-to-one comparison unless you get the japanese version with no translation and the translated one from the same run.
The good things with Whisper end-to-end translation are:
(a) It uses context for translation. It tries to build a context for example guessing gender (he, she), and punctuations for translation task.;
(b) It makes the entire Whisper output faster. Translate tsak is faster than transcribe task. It is funny but their main sw engineer was saying that the way the algorithm is written, the end-to-end trasnlation task is performed faster than just transcribe task 
The good things with DeepL is that It is just a better translator. Fullstop. One bad thing with DeepL is that it often mixes up he/she, it/they, sir/ma'am.For me I decided to just stick with DeepL. I did some comparisons during the early days of Whisper (v1). I haven't done any comparison with v2 but I understand that the translation capability did not change from v1 to v2. To me, DeepL translations came out better. But then again, I don't speek Japanese so my read might be quite wrong.
In terms of being able to compare the outputs as @SamKook suggested, one can make Whisper to be more deterministic by setting both temperature and beam to zero. That makes the output close to determinstic. But the pitfal is that it produces more halucination and repeating lines in the output.
I did one file today with only transcribe and it took nearly 1h 30min instad of approx. 30min, then ran it through Deepl and I wasn't impressed with the result. Did not seem better than just straight letting whisper do the whole job.The good things with Whisper end-to-end translation are:
(a) It uses context for translation. It tries to build a context for example guessing gender (he, she), and punctuations for translation task.;(b) It makes the entire Whisper output faster. Translate tsak is faster than transcribe task. It is funny but their main sw engineer was saying that the way the algorithm is written, the end-to-end trasnlation task is performed faster than just transcribe taskThe good things with DeepL is that It is just a better translator. Fullstop. One bad thing with DeepL is that it often mixes up he/she, it/they, sir/ma'am.
For me I decided to just stick with DeepL. I did some comparisons during the early days of Whisper (v1). I haven't done any comparison with v2 but I understand that the translation capability did not change from v1 to v2. To me, DeepL translations came out better. But then again, I don't speek Japanese so my read might be quite wrong.
In terms of being able to compare the outputs as @SamKook suggested, one can make Whisper to be more deterministic by setting both temperature and beam to zero. That makes the output close to determinstic. But the pitfal is that it produces more halucination and repeating lines in the output.
DeepL is theoretically better, but there's probably some value in doing direct-to-English with the same deep learning model rather than taking the transcribed output and feeding into a second deep learning model that isn't specifically designed to work interact with the first. There's just an additional loss of information during that intermediate step.
It also depends on whether you are using Medium or Large Whisper, and how tuned your parameters are. Some things like an increased Beam Size are going to produce better translations of proper nouns, and Large is just generally better if you can pull it off.