akiba resident JAV subtitlers & subtitle talk★NOT A SUB REQUEST THREAD★
- Thread starter desioner
- 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.
So OpenAI just released Whisper, their speech to text AI, and the transcription seems pretty decent from what I tested using Google Colab.
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
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.
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.
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.
@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 ?
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.
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.@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.@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, @mei for running tests and giving us the data and your experience using this new model that @tangerinefeline was graciously told us all about.
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
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:
thanks for the tip, i'll do that now. I used the small model because it says the large model is for servers.
I decided to test VOSK vs Pytranscriber on the opening of MIAA-698 -[Single Mom Reserve Army] The Hurdles To SEX Are Too Low, Can't Stand My Sweaty Eldest Daughter Naked Every Day! Lima Arai (2022) as the opening is a simple monologue from the father with no music... plus Lima Arai is Sooo hot in this film as the daughter with hyperhydrosis (Over active sweat gland) who is always so hot and sweaty that she never wear cloths (well she wears sock and the tie for her school uniform) that it real deserves subtitles.
I think Pytransriber did a bettre job, what has other peoples experience bee?
| Pytranscriber | Vosk Japanese Large model |
| 1 00:00:00,256 --> 00:00:06,400 Hello. I'm surprised to see an old man out of nowhere. 2 00:00:06,656 --> 00:00:12,800 Please take a few minutes of your time. 3 00:00:13,056 --> 00:00:19,200 This time, I'd like to share with you a little bit about myself. 4 00:00:19,456 --> 00:00:25,600 I'd like to share with you my story. 5 00:00:25,856 --> 00:00:32,000 I'm sorry I'm late to introduce myself. 6 00:00:32,256 --> 00:00:38,400 I'm a little rough around the edges, but I'm the mainstay of my family. I'm right after my father. 7 00:00:38,656 --> 00:00:44,800 As you can see, I'm not a magazine model. 8 00:00:45,056 --> 00:00:51,200 I'm a regular office worker. There's no Araike here. 9 00:00:51,456 --> 00:00:57,600 Because eight years ago, my wife left me for a young man she worked with part-time. 10 00:00:57,856 --> 00:01:04,000 She got tired of him and left him. 11 00:01:04,256 --> 00:01:05,536 Since then, I've been the mainstay of Araike for three years. 12 00:01:05,792 --> 00:01:11,680 I've been the mainstay of Araike's household, raising our three children. 13 00:01:11,936 --> 00:01:14,496 Now then... 14 00:01:15,008 --> 00:01:17,824 I would like to introduce you to my lovely children. 15 00:01:18,336 --> 00:01:24,480 First of all, my son, my eldest, is my favorite, a college student. 16 00:01:24,736 --> 00:01:30,880 He is a serious boy, and I guess he is more solid than I am. 17 00:01:31,136 --> 00:01:33,184 Lately, he's been... 18 00:01:33,440 --> 00:01:35,232 Muscle training, Yamanote... 19 00:01:35,744 --> 00:01:37,536 A typhoon is coming. 20 00:01:37,792 --> 00:01:43,936 Maybe he's got too much power. 21 00:01:44,960 --> 00:01:51,104 Next is Shinji, my second and youngest son. 22 00:01:53,920 --> 00:02:00,064 He doesn't go to school. 23 00:02:00,320 --> 00:02:06,464 Animated? He's always doing that. 24 00:02:06,720 --> 00:02:12,864 A recluse. Me too. 25 00:02:13,120 --> 00:02:17,472 I don't get to see him much. 26 00:02:17,728 --> 00:02:19,520 I don't get to see him much, so I don't know what he's thinking. 27 00:02:19,776 --> 00:02:24,640 I don't know what he's thinking. 28 00:02:25,152 --> 00:02:31,296 I'm Rima, the eldest between two sons. 29 00:02:31,552 --> 00:02:37,696 She's the problem child of Araike. 30 00:02:37,952 --> 00:02:44,096 She never wears clothes and spends most of her time at home completely naked. 31 00:02:44,352 --> 00:02:50,496 She can't even study... and she's not in school. 32 00:02:54,592 --> 00:03:00,736 She's a gal? Also, for some reason, she's always sweaty since she was a kid. 33 00:03:02,528 --> 00:03:05,344 Is she expensive? 34 00:03:05,856 --> 00:03:12,000 This time, Lima is causing all kinds of trouble. 35 00:03:12,256 --> 00:03:18,400 Let's see... naked and covered in sweat with my eldest daughter. | 1 00:00:01,200 --> 00:00:02,400 Hi there. 2 00:00:04,110 --> 00:00:05,520 I'm sure some of you are surprised to see your uncle out of nowhere. 3 00:00:06,180 --> 00:00:08,088 I know some of you were surprised. 4 00:00:09,180 --> 00:00:10,180 It's just for a few hours. 5 00:00:10,680 --> 00:00:11,680 Please bear with me. 6 00:00:13,410 --> 00:00:14,410 This time 7 00:00:14,614 --> 00:00:15,614 I'd like to send you 8 00:00:18,090 --> 00:00:19,680 I don't know if I should say this myself, but... 9 00:00:20,401 --> 00:00:21,401 I'm a little... 10 00:00:23,250 --> 00:00:24,270 I've changed a lot. 11 00:00:25,470 --> 00:00:26,470 My family 12 00:00:27,090 --> 00:00:28,620 It's a story about a coarse house. 13 00:00:30,660 --> 00:00:31,950 I'm late to introduce myself. 14 00:00:33,090 --> 00:00:34,090 I am 15 00:00:34,140 --> 00:00:35,504 The mainstay of our coarse family. 16 00:00:36,240 --> 00:00:37,800 I'm Taku, my father. 17 00:00:38,910 --> 00:00:40,110 As you can see... 18 00:00:40,560 --> 00:00:41,730 I'm not a magazine model 19 00:00:42,270 --> 00:00:43,270 I'm not 20 00:00:45,360 --> 00:00:46,830 I'm just an ordinary office worker 21 00:00:46,830 --> 00:00:47,830 I'm an ordinary businessman. 22 00:00:48,840 --> 00:00:49,840 Washing machine 23 00:00:50,268 --> 00:00:51,268 I don't have a washing machine 24 00:00:52,321 --> 00:00:53,321 I'm not a washer. 25 00:00:53,970 --> 00:00:54,180 8 years ago 26 00:00:54,210 --> 00:00:55,290 years ago, my wife 27 00:00:55,800 --> 00:00:56,070 part time job 28 00:00:56,070 --> 00:00:58,193 I've been working with a lot of young men. 29 00:00:59,670 --> 00:01:00,300 She got tired of me. 30 00:01:00,661 --> 00:01:01,661 She got tired of me. 31 00:01:01,860 --> 00:01:02,860 She left me. 32 00:01:04,140 --> 00:01:05,250 And then I went back to 33 00:01:06,030 --> 00:01:06,750 As the mainstay 34 00:01:06,990 --> 00:01:08,100 As the mainstay 35 00:01:09,120 --> 00:01:09,210 Three... 36 00:01:09,229 --> 00:01:10,229 I've raised three children. 37 00:01:10,410 --> 00:01:11,410 I've raised three children. 38 00:01:13,530 --> 00:01:14,530 And now... 39 00:01:15,000 --> 00:01:17,730 I would like to introduce you to my lovely children. 40 00:01:18,570 --> 00:01:19,920 Let's start from there. 41 00:01:21,630 --> 00:01:22,830 My eldest son Daisuke. 42 00:01:23,340 --> 00:01:24,340 He's a college student, isn't he? 43 00:01:26,100 --> 00:01:27,630 He's very serious. 44 00:01:28,650 --> 00:01:31,200 I guess he's more solid than I am. 45 00:01:31,950 --> 00:01:33,060 What is he doing these days? 46 00:01:33,870 --> 00:01:34,348 muscle training 47 00:01:34,348 --> 00:01:35,348 She's into muscle training. 48 00:01:35,730 --> 00:01:37,440 She seems to be working out a lot. 49 00:01:38,220 --> 00:01:38,591 Maybe he has too much power 50 00:01:38,591 --> 00:01:39,990 Maybe he has too much power. 51 00:01:43,444 --> 00:01:44,640 It might be amazing. 52 00:01:47,370 --> 00:01:48,370 Next... 53 00:01:49,230 --> 00:01:50,230 Second son 54 00:01:50,640 --> 00:01:50,910 Check 55 00:01:50,910 --> 00:01:51,000 of 56 00:01:51,327 --> 00:01:52,327 Check it out! 57 00:01:53,880 --> 00:01:54,960 This boy 58 00:01:56,520 --> 00:01:57,870 He doesn't go to school. 59 00:02:02,400 --> 00:02:03,400 Anime 60 00:02:03,750 --> 00:02:03,990 Mecha 61 00:02:04,230 --> 00:02:04,500 or... 62 00:02:05,130 --> 00:02:06,540 That's all she does. 63 00:02:09,085 --> 00:02:10,085 That's nice. 64 00:02:11,550 --> 00:02:12,550 Me too. 65 00:02:13,230 --> 00:02:15,420 I don't know what he's thinking. 66 00:02:17,790 --> 00:02:19,530 I don't know what he's thinking. 67 00:02:20,220 --> 00:02:21,420 I don't know what he's thinking. 68 00:02:23,880 --> 00:02:24,880 Finally. 69 00:02:25,800 --> 00:02:26,800 The first son 70 00:02:27,180 --> 00:02:28,180 The second son 71 00:02:28,890 --> 00:02:29,890 The eldest daughter. 72 00:02:29,945 --> 00:02:30,945 Now. 73 00:02:32,790 --> 00:02:34,020 This is Koga. 74 00:02:34,650 --> 00:02:36,210 She's the problem child of the family. 75 00:02:36,990 --> 00:02:37,990 First of all... 76 00:02:38,220 --> 00:02:38,640 Clothes. 77 00:02:38,790 --> 00:02:39,790 No clothes. 78 00:02:39,960 --> 00:02:42,480 She spends most of her time at home completely naked. 79 00:02:44,400 --> 00:02:45,400 Also... 80 00:02:45,810 --> 00:02:46,810 I don't know how to say it. 81 00:02:47,430 --> 00:02:48,690 I can't even study. 82 00:02:49,500 --> 00:02:51,390 I didn't finish school. 83 00:02:53,283 --> 00:02:54,283 I'm what's called 84 00:02:54,660 --> 00:02:54,960 Gyaru 85 00:02:54,960 --> 00:02:55,960 I guess you could say 86 00:02:56,790 --> 00:02:57,840 I don't know why. 87 00:02:58,710 --> 00:03:01,200 I've come all this way since I was a kid. 88 00:03:02,550 --> 00:03:04,290 I wonder if she has hyperhidrosis. 89 00:03:06,120 --> 00:03:07,120 This time 90 00:03:07,530 --> 00:03:07,830 This 91 00:03:08,040 --> 00:03:09,040 is 92 00:03:10,800 --> 00:03:12,030 It's going to be a problem. 93 00:03:14,010 --> 00:03:15,207 Well then... 94 00:03:15,750 --> 00:03:16,890 Naked eldest daughter and 95 00:03:18,089 --> 00:03:19,089 covered in |
Attachments
I agree with @avatarthe. Pytranscriber still delivers good results.
I use it mostly for timing but when it translate things correctly than it's a bonus
The sub I'm using for example was tweak by me with Aegisub and Pytranscriber and SubtitleEdit. I have fallen in love with Aeigusb as I used that as my final clean up. IMSCULLY and some of my reddit friends have really shown me the power of it.
As you might guess, this sub is 90% finish but I haven't released it yet.
I use it mostly for timing but when it translate things correctly than it's a bonus
The sub I'm using for example was tweak by me with Aegisub and Pytranscriber and SubtitleEdit. I have fallen in love with Aeigusb as I used that as my final clean up. IMSCULLY and some of my reddit friends have really shown me the power of it.
As you might guess, this sub is 90% finish but I haven't released it yet.
I watched the intro with both subtitles. I could see the difference. VOSK breaks up sentences, but detects more words (not as obvious with this video but others). PyTranscriber have longer sentences. When both are machine translated, its a close tie. Some words were translated better with VOSK, some better with pytranscriber. (For example, pytranscriber translated the second son's name as Shinji, Vosk did not catch it. Vosk translated "anime" "mecha" for second son's interest, but pyTranscriber translated "animated?". Pytranscriber translated about his wife left him for young man 8 years ago, Vosk did not translate that part properly. Vosk said something about washing machines, but Pytranscriber completely missed that. Pytranscriber missed out "I'm Taku, the father") I still prefer VOSK though, mainly because it doesn't miss out on words, and ease of use.
Also I have an issue with pytranscriber not sure if any of you have it. When I transcribe anything longer than 15min or so, it will get timed out between 40%-85%, and get stuck there. I had to split the video in order to transcribe.
This whisper thing is actually AMAZING, much better than vosk or pytranscriber.
Noob question: how do I run it offline without uploading an mp3 everytime?
EDIT: Omg!!! I watched it with whisper sub, it's amazing!!!! about 90% accuracy compare to vosk/pytranscriber 50%.
Last edited:
Okay, I'll give away the dirty little secret of how to get Pytranscriber to work flawlessly and no crash. I've not waned this to get out because if too many people start using it it could stop working...
SO the big secret it.... Use a VPN witha Japanese address when you run Pytranscriber, that it! It will work flawlessly!
HAHA ok. Anyways, look at my post above. I think you no longer want to use pytranscriber anymore.
@SUNBO would you share the subs from Whisper for MIAA-698.
@avatarthe would you share the subs from pyTranscriber and Vosk for MIAA-698.
This is a very good test case.
Thanks both !