akiba resident JAV subtitlers & subtitle talk★NOT A SUB REQUEST THREAD★

TreeAnt

New Member
May 2, 2008
9
21
This is not 100% accurate and I used to believe that as well, before I started tweaking the various tools that I use now.
If there was a command-line utility I could feed the audio through in a filtering + conversion process I could totally implement it into my scripts. I suspect that Google's Speech-to-Text API already does filtering on their end, but it wouldn't hurt giving it a try and do an A-B test to see if passing the audio through some program yielded better results.
When I see that even names aren't translated properly with auto subs ... I saw a guy who started a streaming website with uncensored FC2 vids (the ones already free everywhere) with auto subs. I had a good laugh as, as I said to him, it didn't make any sense (even very simple lines), like full random words.
I've been impressed by some of the results, and plenty disappointed as well. It's really a mixed bag. $2 auto subs versus what would probably cost hundreds of dollars for a human subtitle job (and then auto-translating because... well... it's typically "good enough" for personal consumption) is what I'm stuck with. Even so, it's still a lot of fun!

If I won a lottery and suddenly had millions I'd love commissioning professional subtitling work, don't doubt it! xD
 
  • Like
Reactions: shinjixx and Taako

Taako

Well-Known Member
May 25, 2017
586
441
When I see that even names aren't translated properly with auto subs ... I saw a guy who started a streaming website with uncensored FC2 vids (the ones already free everywhere) with auto subs. I had a good laugh as, as I said to him, it didn't make any sense (even very simple lines), like full random words.


Actually, there are. My service team has japanese members (even women).
No harm. I said a few, but I feel it should be higher given the income potential imo :D
 

Taako

Well-Known Member
May 25, 2017
586
441
$2 auto subs versus what would probably cost hundreds of dollars for a human subtitle job (and then auto-translating because... well... it's typically "good enough" for personal consumption) is what I'm stuck with. Even so, it's still a lot of fun!

If I won a lottery and suddenly had millions I'd love commissioning professional subtitling work, don't doubt it! xD
And that's me in nutshell.
Although if I won the lottery, I'm paying exclusively for AIKA and some of my favorite JAV actresses to satisfy my needs in ways a video can not. :p :p :p

Anyways, I've seen some great subs and then some subpar ones. Some have changed very little of the dialog and other have changed a lot.

The excuses can be the subtitler didn't know what's being said OR they found the actual dialog "boring".

It's all fun, BUT incredibly frustrating.
 
Last edited:
  • Like
Reactions: mei2

norton12

New Member
Sep 6, 2013
1
0
can I have ROYD-068 in SRT ver.?

I found the Chinese one in Avgle but it was already hard coded into the movie.

I tried Subtitlecat and google. But couldn't find an English translation or an SRT file.
 

Attachments

  • ROYD-068.jpg
    ROYD-068.jpg
    169.9 KB · Views: 26

maload

Active Member
Jul 1, 2008
410
37
When I see that even names aren't translated properly with auto subs ... I saw a guy who started a streaming website with uncensored FC2 vids (the ones already free everywhere) with auto subs. I had a good laugh as, as I said to him, it didn't make any sense (even very simple lines), like full random words.


Actually, there are. My service team has japanese members (even women).


today i search on the net and i found " avsubtil****** "
o great
i have been clicked on that and i watched movies .
no sub at all . its just a movie.
the worst i have seen in this business
at least he should give me something for reading when i am fapping
 

darksider59

Akiba Citizen
Feb 24, 2014
2,366
1,645
can I have ROYD-068 in SRT ver.?

I found the Chinese one in Avgle but it was already hard coded into the movie.

I tried Subtitlecat and google. But couldn't find an English translation or an SRT file.
In the topic title > ★NOT A SUB REQUEST THREAD★

today i search on the net and i found " avsubtil****** "
o great
i have been clicked on that and i watched movies .
no sub at all . its just a movie.
the worst i have seen in this business
at least he should give me something for reading when i am fapping
All their subs are auto translated in english and few other languages.
 
  • Like
Reactions: Imscully

Taako

Well-Known Member
May 25, 2017
586
441
If there was a command-line utility I could feed the audio through in a filtering + conversion process I could totally implement it into my scripts. I suspect that Google's Speech-to-Text API already does filtering on their end, but it wouldn't hurt giving it a try and do an A-B test to see if passing the audio through some program yielded better results.
Wouldn't some movies with loud music stop this process?

I swear whoever edits some of these movies need to know basic post production 101:mad: The music should NEVER be louder than the dialog.
 

ssjgoku4

Member
Jan 20, 2021
27
38
I'm a software engineer with nearly a decade of experience. Today I figured out how to use Google Cloud Compute to transcribe and translate JAVs pretty reliably using some engineer's open source project on GitHub I found.

Rough estimates but...
From start to finish it takes me about 90 seconds of manual manipulation and perhaps 10 minutes of waiting to generate an SRT file.

If I batch translate multiple at once I can done one after the other after the other.

I'm still on Google Compute Cloud's free trial lasting 90 days and offering $300 of credit, but by my rough estimate it costs maybe $0.10-0.25 USD in API costs to translate a 2 hour video. Totally pulling that number out of my ass but the point is it's cheap.

Hope to record and show a demo of this soon. I know I have like zero presence on this Forum but I've had this account for a long time and built most of my JAV collection here. So... feel like sharing in some of the amazing good fortune.

Update 1 - 2022-09-10
It's late and I got way too into this tool. I translated about 70-80 hours or so of JAV material and it cost me about $21.80 in credits. That averages about $0.30 USD an hour. So, my estimate was off, but it's still pretty cheap. Again, I've got $300 of complimentary credit. So whatever.

I might not be able to record a demo of the thing working for a couple of days (I realize I'm going to need to do some editing to hide persona details from a screen recording I hope to make). I have a day job. But I hope to share some updates on how I might translate for people with cool JAVs. I'll give details, but with a free tool on PC/MacOS/Linux anyone should be able to get me the material to perform a translation with a simple share of an audio-only mkv that takes mere moments to create and will only be a like 100mb to send to me.

I'll figure out a method to share the SRTs I've created thus far.

Update 2 - 2022-09-10
These are all entirely random. And most of them are probably closer to a decade old than not. What can I say? I don't download JAVs as often as I used to and I'm pretty content with the library I built in my early 20s. Lots of incest and kinky stuff.

If someone viewing this happens to potentially have one of these videos, give it a try!

Even if you don't download a copy anyway.

For any engineers out there. Have at it: TreeAntSan/trans_video_subs: Auto Translate and add subtitle to video on Google Cloud (github.com)

This is seriously advanced stuff. So... uhhh... not going to trouble shoot your issues, so don't even ask. Wait until I reply again.

Update 3 - 2022-09-10
Haha, holy shit. Never mind about the price prediction. My credit count got updated. It's looking like I'm getting charged about a $1 USD an hour. Still cheap. But this evening of rallying through a bunch of JAVs cost $80 USD. Hoo boy. Still pretty worth it.

Gotta say, top of all. Only going to get fun results with JAVs with roleplay or story. There needs to be a lot of talking. Otherwise you're liable to pay good money for just "Ohh it's so big" or "don't stop"... which isn't fun.

Update 4 - 2022-09-11
More SRT files shared. Now 39 in total (7 more than yesterday).

I've added a slew of bash scripts (scripts you can use in Linux and MacOS) that remove most of the repetitive tasks that'd otherwise require manipulation with MKVToolNix GUI.

Update 5 - 2022-09-11
Oh, and I also discovered a way to cut my Speech-to-Text API costs (the greatest source of cost) down by about 66%. If you let Google perform analysis on your uploaded audio they give you a 33% discount. And they also charge per channel. So a stereo audio track doubles the cost. So, I obviously adjusted the tools to mix down all audio to mono. That's makes the API 66.67% cheaper to use.

From $0.012 USD per 15 seconds to $0.004 USD per 15 seconds. That's equivalent to dropping a 2hr JAV from costing $5.76 USD ((2*60*60/15) * 0.012) to $1.92 USD ((2*60*60/15) * 0.004).

I'll have to trust the price drop will come, I'm only fairly certain I've dropped the price 66%.

Next comes translation. While each transcript's length varies, I imagine we can expect anywhere from 7,000 to 15,000 characters to be used per video (this is a big guess-timate, I don't currently have any samples of actual counts, really just doing word-counts of English after-the-fact).

Translate API costs come to $20 USD per 1,000,000 characters translated (incoming). It helps that Japanese consumes fewer characters than English. So if a video had 10,000 Japanese characters that'd be a pretty humble $0.20 USD.

So, a two hour video would probably cost $2.12 USD or so to complete. I checked the cost of the Cloud Run instance and it's really quite small. In generating 39 videos' worth of translations, the cost of the computing has only come to $5.69 USD so far, which is tiny.
I created my own method too using mozilla deepspeech, i collected all the Japanese speech database i could find to train the AI and after many test and trails i got astonishing result. But due to the complication of using it i didn't share it here as it will go over the head of average users. I am working on making a user friendly version will release if it gets completed.
 
  • Like
Reactions: pspo and Imscully

maload

Active Member
Jul 1, 2008
410
37
I created my own method too using mozilla deepspeech, i collected all the Japanese speech database i could find to train the AI and after many test and trails i got astonishing result. But due to the complication of using it i didn't share it here as it will go over the head of average users. I am working on making a user friendly version will release if it gets completed.
I created my own method too using mozilla deepspeech, i collected all the Japanese speech database i could find to train the AI and after many test and trails i got astonishing result. But due to the complication of using it i didn't share it here as it will go over the head of average users. I am working on making a user friendly version will release if it gets completed
i think you talk about you create japanese model ?
can you share your japanese pretrained model


 
  • Like
Reactions: mei2

tangerinefeline

New Member
Sep 22, 2022
3
7
So OpenAI just released Whisper, their speech to text AI, and the transcription seems pretty decent from what I tested using Google Colab.

Screenshot 2022-09-22 17.34.33.png

It's also relatively simple since it's only one command to do a Japanese -> English transcription.

Here's the Github repo for it: https://github.com/openai/whisper

Here's the Colab notebook for it, just replace the file_location variable with a link to the audio file you want to transcribe: https://colab.research.google.com/drive/1j3-_EF43nUCeIkrzpk_jpamtFZmURYrU?usp=sharing
 

maload

Active Member
Jul 1, 2008
410
37
So OpenAI just released Whisper, their speech to text AI, and the transcription seems pretty decent from what I tested using Google Colab.

View attachment 3045996

It's also relatively simple since it's only one command to do a Japanese -> English transcription.

Here's the Github repo for it: https://github.com/openai/whisper

Here's the Colab notebook for it, just replace the file_location variable with a link to the audio file you want to transcribe: https://colab.research.google.com/drive/1j3-_EF43nUCeIkrzpk_jpamtFZmURYrU?usp=sharing
thank you so much , i want to test it .
 
  • Like
Reactions: tangerinefeline

Taako

Well-Known Member
May 25, 2017
586
441
So OpenAI just released Whisper, their speech to text AI, and the transcription seems pretty decent from what I tested using Google Colab.

View attachment 3045996

It's also relatively simple since it's only one command to do a Japanese -> English transcription.

Here's the Github repo for it: https://github.com/openai/whisper

Here's the Colab notebook for it, just replace the file_location variable with a link to the audio file you want to transcribe: https://colab.research.google.com/drive/1j3-_EF43nUCeIkrzpk_jpamtFZmURYrU?usp=sharing
Thank you for sharing.
Have you tested during scenes that are low speaking, multiple people talking, and background noises such as music and etc?
Those are the common complaints and should be the real test.
 

tangerinefeline

New Member
Sep 22, 2022
3
7
Thank you for sharing.
Have you tested during scenes that are low speaking, multiple people talking, and background noises such as music and etc?
Those are the common complaints and should be the real test.

Hmm not yet, to be honest I usually only lurk here, is there a human made translation that would be good to do a comparison to? I think this is useful when you have videos that might be a little bit older or obscure which are difficult to find subtitles for, and definitely not a replacement yet for human transcription.
 
  • Like
Reactions: Taako

Taako

Well-Known Member
May 25, 2017
586
441
Hmm not yet, to be honest I usually only lurk here, is there a human made translation that would be good to do a comparison to? I think this is useful when you have videos that might be a little bit older or obscure which are difficult to find subtitles for, and definitely not a replacement yet for human transcription.
I appreciate your honest answer.
I'm not surprise you mention older jav. I regularly subbed older jav because no one really created subs for them and I think the movies are better acted and have more convincing sex scenes. :D

Also many old jav don't have the loud background music and multiple people talking. It's good for subbing.

It seem all the models work generously given clear speaking and less background noise. I'll be impress when a translation software can do that and be affordable. I'm sure many govt security, police forces, military, and movie studios have and uses the tech.
 

mei2

Active Member
Dec 6, 2018
70
132
So OpenAI just released Whisper, their speech to text AI, and the transcription seems pretty decent from what I tested using Google Colab.

View attachment 3045996

It's also relatively simple since it's only one command to do a Japanese -> English transcription.

Here's the Github repo for it: https://github.com/openai/whisper

Here's the Colab notebook for it, just replace the file_location variable with a link to the audio file you want to transcribe: https://colab.research.google.com/drive/1j3-_EF43nUCeIkrzpk_jpamtFZmURYrU?usp=sharing

@tangerinefeline , thanks for this. New tools and methods are always welcome!

The technology (and the research paper) look quite promising. So I quickly ran few tests on it --thanks for the link to the colab. Makes things so much faster. Beside, the large models need like 16GB VRAM :).

I ran Whisper on 2 new files released today: SSIS-525 Aoi Tsukasa, And JUQ-098 Nao Jinguuji. I ran default model w translation, default model, large model, and medium. FWIW, here is what I observed so far:

- Overal: it is quite promising but it seems that at least the Japanese model needs more training and calibration.
- Ease of use: Good! The colab especially makes it very easy to use.
- Dialogue detection: Not very good.
- Timing: Not good. Transcription is all over the place. I have a feeling that this is caused by Huggingface transformers.
- Translation: I suggest not to use it. The one version that I looked at seemed to be off.
- Comparison with VOX: the vox version through subtitleedit produces much more accurate dialogue timing. But the dialogue detection seem to be in par with Whisper.


I'm keen to hear from @ssjgoku4 about his retrained model. We seem to have some quite tech-smart people in this forum. It would be great if we put our efforts together to work on training (a selected) model for JAV. May be we can setup a Patreaon perhaps ?
 

Attachments

  • JUQ-098.ja Nao Jinguuji - WHISPER MEDIUM.zip
    18 KB · Views: 47
  • SSIS-525.ja Aoi Tsukasa - WHISPER LARGE.zip
    1 KB · Views: 26

SUNBO

Member
Nov 19, 2007
63
36
can I have ROYD-068 in SRT ver.?

I found the Chinese one in Avgle but it was already hard coded into the movie.

I tried Subtitlecat and google. But couldn't find an English translation or an SRT file.
If you are desperate you can use your phone's app "google translate" and use the camera option. But the problem with this is.... you need a third hand. One hand to hold the phone, one hand to skip forward, and one hand to touch your little brother.
 

SUNBO

Member
Nov 19, 2007
63
36
@tangerinefeline , thanks for this. New tools and methods are always welcome!

The technology (and the research paper) look quite promising. So I quickly ran few tests on it --thanks for the link to the colab. Makes things so much faster. Beside, the large models need like 16GB VRAM :).

I ran Whisper on 2 new files released today: SSIS-525 Aoi Tsukasa, And JUQ-098 Nao Jinguuji. I ran default model w translation, default model, large model, and medium. FWIW, here is what I observed so far:

- Overal: it is quite promising but it seems that at least the Japanese model needs more training and calibration.
- Ease of use: Good! The colab especially makes it very easy to use.
- Dialogue detection: Not very good.
- Timing: Not good. Transcription is all over the place. I have a feeling that this is caused by Huggingface transformers.
- Translation: I suggest not to use it. The one version that I looked at seemed to be off.
- Comparison with VOX: the vox version through subtitleedit produces much more accurate dialogue timing. But the dialogue detection seem to be in par with Whisper.


I'm keen to hear from @ssjgoku4 about his retrained model. We seem to have some quite tech-smart people in this forum. It would be great if we put our efforts together to work on training (a selected) model for JAV. May be we can setup a Patreaon perhaps ?
Hey thanks for mentioning that SubtitleEdit can transcribe audio! It seems to be available only from March 2022 this year. I just tested it on one of my favourite JAV (KTFT-004) and had great results. 95% of it were transcribed, 70% were translated accurately (meaning it made sense). Which was a huge improvement from pyTranscriber which I had issues with, only about 30-50% transcribed due to it using google speech recognition API and google API is bad at detecting voice that's a bit distant from camera.
 
  • Like
Reactions: mei2

Taako

Well-Known Member
May 25, 2017
586
441
@tangerinefeline , thanks for this. New tools and methods are always welcome!

The technology (and the research paper) look quite promising. So I quickly ran few tests on it --thanks for the link to the colab. Makes things so much faster. Beside, the large models need like 16GB VRAM :).

I ran Whisper on 2 new files released today: SSIS-525 Aoi Tsukasa, And JUQ-098 Nao Jinguuji. I ran default model w translation, default model, large model, and medium. FWIW, here is what I observed so far:

- Overal: it is quite promising but it seems that at least the Japanese model needs more training and calibration.
- Ease of use: Good! The colab especially makes it very easy to use.
- Dialogue detection: Not very good.
- Timing: Not good. Transcription is all over the place. I have a feeling that this is caused by Huggingface transformers.
- Translation: I suggest not to use it. The one version that I looked at seemed to be off.
- Comparison with VOX: the vox version through subtitleedit produces much more accurate dialogue timing. But the dialogue detection seem to be in par with Whisper.


I'm keen to hear from @ssjgoku4 about his retrained model. We seem to have some quite tech-smart people in this forum. It would be great if we put our efforts together to work on training (a selected) model for JAV. May be we can setup a Patreaon perhaps ?
Thank you both for testing this new tech and capabilities.

Thank you, @mei for running tests and giving us the data and your experience using this new model that @tangerinefeline was graciously told us all about. :D

And @mei, i for one appreciate you taking the time to check out this new translation and giving us a through analyst and your thoughts on it.

It's classy acts like this is why I really like AKIBA Online, the people and the moderators are just so much down to earth, fun, and professional:cheers:
 

mei2

Active Member
Dec 6, 2018
70
132
Hey thanks for mentioning that SubtitleEdit can transcribe audio! It seems to be available only from March 2022 this year. I just tested it on one of my favourite JAV (KTFT-004) and had great results. 95% of it were transcribed, 70% were translated accurately (meaning it made sense). Which was a huge improvement from pyTranscriber which I had issues with, only about 30-50% transcribed due to it using google speech recognition API and google API is bad at detecting voice that's a bit distant from camera.

If you haven't done it yet, download the Vosk Japanese Large model --it produces better results for me than the default Japanese Small model.
 
Last edited: