Post your JAV subtitle files here - JAV Subtitle Repository (JSP)★NOT A SUB REQUEST THREAD★

panop857

Active Member
Sep 11, 2011
155
222
Right, I tried my first whisper translation. I got a lot of repeating lines, but not just a single line repeating (though there were some of those as well) but groups of lines. For example:

123 00:08:16,000 --> 00:08:18,000 Is it true?
124 00:08:18,000 --> 00:08:22,000 No, it's not.
125 00:08:22,000 --> 00:08:25,000 I'm still in the hospital.
126 00:08:25,000 --> 00:08:28,000 I'm still in the hospital.
127 00:08:28,000 --> 00:08:30,000 Is it true?
128 00:08:30,000 --> 00:08:32,000 No, it's not.

Does anyone have an idea what settings I should change to avoid these?

I am running it on my own computer instead of online, because the online thing kept giving me errors. This means I have to use the small model, I hope that's not what is causing this issue.

Set "condition_on_previous_text" or whatever it is to False. In theory, it is there to make it learn into consistent style and slightly bias the results towards specific topics that the entire source is about. In practice specifically for porn films of languages other than English, there's going to be a lot of scenes where it can't really make any good guess, which leads to the slight bias on previous text pushing it towards the exact same "answer", and then that bias for the previous bias layers with the bias from two lines ago, and so on, leading it to land in long ruts of the same sentence it it gets a nice clear answer for it to change its mind.

The Small model definitely contributes to this, but the condition_on_previous_text is probably the larger influence.

The Small model will be enough to get the general gist of plot scenes, but will do much worse on sex dialogue and will need quite a bit more editing and interpretations to get into a good spot. If you are limited, it can be good practice to run the same film through several different settings, making sure to save the outputs to different names, and reading them side by side and copying the best into a new file.
 

Electromog

Akiba Citizen
Dec 7, 2009
4,436
2,705
I'm less interested in the dialog during sex (as it is generally much the same anyway) but more in the dialog of the story surrounding the sex, so that's ok. However, github claims 8 GB of video memory should be enough for the medium model, so maybe I should try that next.

I will try the condition_on_previous_text option. I guess since it only takes a few minutes I can do a few runs with different settings and see how that affects things.
 

panop857

Active Member
Sep 11, 2011
155
222
I'm less interested in the dialog during sex (as it is generally much the same anyway) but more in the dialog of the story surrounding the sex, so that's ok. However, github claims 8 GB of video memory should be enough for the medium model, so maybe I should try that next.

I will try the condition_on_previous_text option. I guess since it only takes a few minutes I can do a few runs with different settings and see how that affects things.

If you're trying to get away with just Small, and it only takes a few minutes, you can try to do runs with a low temperature (but not zero) and like best_of 10 or something. Part of the key to Whisper is doing settings that try to compensate for whatever your computational stuff is limited by.
 

Makkdom

Well-Known Member
Mar 4, 2019
157
385

APNS-305 As long as it's for the sake of the person I love, I don't mind being circulated, defiled, or seeded by people I don't know... Jun Suehiro​


Another file I found and worked on for a while. A young girl agrees to help her boyfriend pay off some debts and spirals into a pretty degrading life.


apns00305pl.jpg
 

Attachments

  • APNS-305-en.zip
    15 KB · Views: 313
Last edited:

SamKook

Grand Wizard
Staff member
Super Moderator
Uploader
May 10, 2009
3,536
4,897
Whenever I use the Subtitle Edit program I always like to convert to Subrip type file once the auto translate function is finished. Not even sure how I landed on that as a fav. Perhaps I'm missing out, but it seems to streamline the subs while maintaining the ability to edit properly. I'd be curious to know if there is a reason you prefer ASS (as opposed to Subrip or ASSA).

ASSA is the same thing as .ass btw.

Unless you want to do something more than just basic text on a video, there's really no good reason to use .ass over subrip(.srt) which is likely the most supported subtitle format.

.ass just gives you a lot of control over how the text is displayed. You can change the font of individual lines, color them, move them to a specific position, scale them bigger or smaller, add gradients to the text, fade effects, stuff like that. You can even draw shapes if you want to block some text on screen and add subtitles over that part.

Here's an example of something fancy I did back in the day, the "B&L's Style Romance!" text in that picture is 100% soft sub, you can toggle it on and off. It's 9 different subtitle lines(and possibly 2 more lines for the gradient and something else, forgot how that works 7 of which are to create the gradient effect, it only shows a slice of each line of different colors) stacked onto each other to mimic the style of the japanese title.
MM08_title.jpg

Edit: One thing I forgot about that could be useful with .ass even if you only want basic text is that you can set an actor name for each line which doesn't get displayed into the subs so if you find it helpful to know who is saying what when editing, that's one thing .srt can't do.
 
Last edited:

ironfevers

Active Member
Nov 9, 2019
37
157
Just a tip on Whisper about translation: Don't use it. Just choose No Translation and do that later in Google translate or DeepL translate

I have experimented with VAD-Threshold and have settled on 0.3. What does Chunk_Threshold (3.0) do? Size of the audio parts analyzed? I'll try to lower and raise the number to see if there are improvements. Source separation didn't work for me. I get error messages.
I also arrived at the same conclusions. After extensive testing, I've settled on 3.2 chunk_threshold and 0.3 vad_threshold for a good balance between catching short sentences and natural sounding speech. Source separation also fails to work. The spleeter package is to blame.
I have increased volume with all three types of audio to text and if you don't go overboard with the volume increase it has always yielded better results. Amplification is also better than compression. Some software doesn't like clipping so you may have to think about that, too.
I also use amplification. I amplify 1 db in Audacity and export as wav.


I have Whisper with VAD installed locally for long duration videos such as HUNTB, but I use the medium model with an RTX 3070 as large requires more VRAM. I also tested max_retries. One file gave me 256 chunks. 52 chunks (20%) needed a retry. Of those 52 chunks, 45 (87%) still failed with a max_attempts of 24. The rest succeeded within 3 retries. Conclusion, stick with a max_attempts value of 1, 2 or 3. I stick with no translation and just translate using DeepL. No one talked about the garbage_list. It'll save you so much time from cleaning up the subtitle. I added over 1000 Japanese words to the garbage list. https://drive.google.com/file/d/1AwJ-VdxA4jM6yJS37Tdxrbj1uAgaU08j/view?usp=share_link. Anything that sounds like laughing, screaming, moaning, etc. get deleted from the subtitle. The garbage list is not complete, there is always more to add. Also I changed the out_path filename so that it includes the settings used. No need to upload to the sidebar in Colab. If mounting your google drive, an example audio_path is "/content/drive/My Drive/temp/abc-123-1db.wav".

aldn034pl.jpg
gvh454pl.jpg
 

Attachments

  • gvh-472-large.whisper.ja.zip
    27.1 KB · Views: 306
  • huntb-241-medium.whisper.ja.zip
    107.7 KB · Views: 279
  • huntb-222-medium.whisper.ja.zip
    55.3 KB · Views: 294
  • gvh-454-large.whisper.ja.zip
    51.8 KB · Views: 268
  • aldn-034-medium.whisper.ja.zip
    78 KB · Views: 287
  • gvh472pl.jpg
    gvh472pl.jpg
    179.9 KB · Views: 227
  • huntb241pl.jpg
    huntb241pl.jpg
    209.2 KB · Views: 229
  • huntb222pl.jpg
    huntb222pl.jpg
    213.7 KB · Views: 225

Chuckie100

Well-Known Member
Sep 13, 2019
505
1,963
RCTD-505 If You Can Endure Ejaculation With Your Mother's Doskebe Fellatio,
A Prize Of 1 Million Yen If You Explode Incest Punishment Game 2
342742_3xplanet_RCTD-505_cover.jpg

I used Whisper to produce this subtitle file for RCTD-505. This is the latest mother-son JAV from Rocket. 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.​

 

Attachments

  • RCTD-505 - Copy.rar
    23 KB · Views: 347

r00g

Member
Jul 25, 2009
84
70
@ironfevers - thanks for the Whisper-VAD.py file! I started to do the same thing, but went a different direction after some thrashing around w/ Python. Will give this a shot soon.
 

Imscully

Well-Known Member
Apr 1, 2014
293
500
ASSA is the same thing as .ass btw.

Unless you want to do something more than just basic text on a video, there's really no good reason to use .ass over subrip(.srt) which is likely the most supported subtitle format.

.ass just gives you a lot of control over how the text is displayed. You can change the font of individual lines, color them, move them to a specific position, scale them bigger or smaller, add gradients to the text, fade effects, stuff like that. You can even draw shapes if you want to block some text on screen and add subtitles over that part.

Here's an example of something fancy I did back in the day, the "B&L's Style Romance!" text in that picture is 100% soft sub, you can toggle it on and off. It's 9 different subtitle lines(and possibly 2 more lines for the gradient and something else, forgot how that works 7 of which are to create the gradient effect, it only shows a slice of each line of different colors) stacked onto each other to mimic the style of the japanese title.
View attachment 3164793

Edit: One thing I forgot about that could be useful with .ass even if you only want basic text is that you can set an actor name for each line which doesn't get displayed into the subs so if you find it helpful to know who is saying what when editing, that's one thing .srt can't do.
Thanks for the great, comprehensive response.
Much appreciated.
I owe you.
 
  • Like
Reactions: Taako

Chuckie100

Well-Known Member
Sep 13, 2019
505
1,963

DTKM-030 And Because I’ll Not Inspire My Mother, And Me Yarra Let Kimi Mom. Wada HyakuMika Okumura Pupil

1avnsuorhjh0.jpg

I used Whisper to produce this subtitle file for DTKM-30. This is another mother-son swap JAV, staring two hot MILFs and where the son gets busted at the end! I had to keep updating the names as Whisper was not consistent. Still not sure if I got them totally correct! 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.​

 

Attachments

  • DTKM-030A.rar
    14.3 KB · Views: 202
  • DTKM-30B.rar
    9.1 KB · Views: 186

Chuckie100

Well-Known Member
Sep 13, 2019
505
1,963

JUY-887 One Weekend I Lied To My Wife About Going On A Business Trip, And Once She Thought I Had Left, She Fucked The Shit Out Of The Guy She's Been Cheating With. I Stayed Hidden In The House The Whole Time And Saw Everything... Eriko Miura

juy887pl.jpg

I used Whisper to produce this subtitle file for JUY-886 starring Eriko Miura. As always however, I still had to clean it up a bit and re-interpreted some of the meaningless/ "lewd-less" dialog. And Yea, I twisted it to be about a mother and her son! (You can untwist it be replacing "...father" with "...husband" and appropriate pronoun.) Again, I don't understand Japanese or Chinese so my other 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.​

 

Attachments

  • JUY-887.rar
    12.6 KB · Views: 244

Imscully

Well-Known Member
Apr 1, 2014
293
500

JUY-887 One Weekend I Lied To My Wife About Going On A Business Trip, And Once She Thought I Had Left, She Fucked The Shit Out Of The Guy She's Been Cheating With. I Stayed Hidden In The House The Whole Time And Saw Everything... Eriko Miura

juy887pl.jpg

I used Whisper to produce this subtitle file for JUY-886 starring Eriko Miura. As always however, I still had to clean it up a bit and re-interpreted some of the meaningless/ "lewd-less" dialog. And Yea, I twisted it to be about a mother and her son! (You can untwist it be replacing "...father" with "...husband" and appropriate pronoun.) Again, I don't understand Japanese or Chinese so my other 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.​

Thanks, Chuckie!!!
 
  • Like
Reactions: Taako

Safadinho

New Member
Dec 26, 2021
15
9

DTKM-030 And Because I’ll Not Inspire My Mother, And Me Yarra Let Kimi Mom. Wada HyakuMika Okumura Pupil

1avnsuorhjh0.jpg

I used Whisper to produce this subtitle file for DTKM-30. This is another mother-son swap JAV, staring two hot MILFs and where the son gets busted at the end! I had to keep updating the names as Whisper was not consistent. Still not sure if I got them totally correct! 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.​

Thank you very much for this. This looks really good.
 
Last edited:
  • Like
Reactions: Taako and Imscully
Feb 25, 2023
13
39
I purposefully included the address bar in the screenshot from the post I made above yours so people would know, but here it is anyway, it's the WhisperWithVAD link people keep posting here: https://colab.research.google.com/github/ANonEntity/WhisperWithVAD/blob/main/WhisperWithVAD.ipynb


And yeah, seems to be that way looking at the github page for the cpp port someone made, it probably doesn't include the internal ffmpeg processing yet that the python version has:
Code:
Note that the main example currently runs only with 16-bit WAV files, so make sure to convert your input before running the tool. For example, you can use ffmpeg like this:
ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav
You should also make it mono(-ac 1) unlike your previous example or else you're getting a file that's double the size for nothing.
I know this is a little old, but I've been running whisper on an mac M1. I experimented around with the different models and found the medium model to perform slightly better. Though the best translation I've found was from a fine-tuned model on hugging face that was specific for japanese. I'm pretty impressed with combining whisper (transcription) and deepl (translation).
 
Feb 25, 2023
13
39
I also arrived at the same conclusions. After extensive testing, I've settled on 3.2 chunk_threshold and 0.3 vad_threshold for a good balance between catching short sentences and natural sounding speech. Source separation also fails to work. The spleeter package is to blame.

I also use amplification. I amplify 1 db in Audacity and export as wav.


I have Whisper with VAD installed locally for long duration videos such as HUNTB, but I use the medium model with an RTX 3070 as large requires more VRAM. I also tested max_retries. One file gave me 256 chunks. 52 chunks (20%) needed a retry. Of those 52 chunks, 45 (87%) still failed with a max_attempts of 24. The rest succeeded within 3 retries. Conclusion, stick with a max_attempts value of 1, 2 or 3. I stick with no translation and just translate using DeepL. No one talked about the garbage_list. It'll save you so much time from cleaning up the subtitle. I added over 1000 Japanese words to the garbage list. https://drive.google.com/file/d/1AwJ-VdxA4jM6yJS37Tdxrbj1uAgaU08j/view?usp=share_link. Anything that sounds like laughing, screaming, moaning, etc. get deleted from the subtitle. The garbage list is not complete, there is always more to add. Also I changed the out_path filename so that it includes the settings used. No need to upload to the sidebar in Colab. If mounting your google drive, an example audio_path is "/content/drive/My Drive/temp/abc-123-1db.wav".

View attachment 3164884
View attachment 3164885
So I do something similar I have one list that are just strings to ignore. A lot of 'door closing', 'ghost screaming' or various sounds. I also have a two column file that i use for minor replacements, but yeh the file keeps getting larger and larger. I'll do an initial translation, and look at all the lines sorted in a text file and then copy the bad lines over to my text files i use to filter. I've noticed a lot of my bad lines have ()s in them so makes it a little easier. I use the medium model as well, not because of speed, but just because I noticed it tends to be more accurate. After processing with deepl I'll have 3 SRT files, one english, one japanese and one i leave as dual.
 

TmpGuy

JavLuv author, lesbian connoisseur
Jun 1, 2013
609
769
Here's another update of my "all subtitles ever posted to this thread" collection.

Since then, there have been 722 subtitles posted.
v4-akiba-import.png

Version 4 is now up to 33,066 subtitles. All files have been renamed, language IDs added, checked for issues, and sorted alphabetically by ID. Enjoy!

FileJoker:

Rapidgator:

Quick plug: JavLuv 1.1.22 was used to sort these subtitles, and can also match your movie collection to available subtitles as well. It's completely free and open source, and I support it here in these forums.
 

ericf

Well-Known Member
Jan 13, 2007
234
528
I've tried both stereo and mono and from what I can tell, I always get better results starting with a stereo file. It makes no difference if Whisper converts that stereo track into a mono track. My tests point to better results if the original track is in stereo. And for me there's pretty much no difference between two passes on the same file. I might get a couple more commas or full stops and one or two words in a few sentences will be different, but on the whole, I get the same content. If I change the Threshold, though, there are always clear differences. So, for me, I'll stick with my stereo uploads.