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

Thanks to bestbuy for his info on SubtitleEdit and the inclusion of FasterWhisperXXL capabilities. I ran WhisperColab, The SubtitleEdit FasterWhisperXXL and the FasterWhisperXXL from the Microsoft Store (that I had purchased), using RCT-746 as a test case. My initial assessment is that WhisperColab might be a little better. Both of the FasterWhisperXXL produced Subs appear to be pretty similar and could be used to fill in missing dialog and provide clarification. I realize this is a single data point and may not represent all cases but I think I will continue my subscription to WhisperColab and use the SubtitleEdit FasterWhisperXXLto help in the editing (while also continuing evaluation of FasterWhisper. Anyway I'd love to hear about other's experiences with FasterWhisperXX and what your assessment of FasterWhisperXXL is.

One last thing: FasterWhisperXXL is indeed faster than using WhisperColab, mainly I think because one does not have to upload an audio file.
 

Attachments

Hi Chuckie! Just wanted to drop by and say thanks for the subs you and others have posted here in this thread! Love it! I've been on a huge Hana Haruna binge lately and this series with her nephew is gold! She has so many delicious mom titles out there as well. Again thanks a bunch!
Yea, I'd love to "binge" on Hana's body too! ;)
 
  • Like
Reactions: DapDunlap

[Reducing Mosaic]RCT-746 Target Wary Mom Chibikko Sexual Harassment Molester Corps Bathing Hen Request SP​

1rct746pl.jpg

I recently downloaded this reduced mosaic of an oldie Mother-Son themed JAV movie. I used Whisper Colab to create this Sub and I also used a Sub created with FasterWhisperXXL that I used to fill in missing dialog and provide better clarification in several places...and oh yea, I enhanced it a tad! Finally, I attempted to clean it up a bit and re-interpreted some of the meaningless/ "lewd-less" dialog. Again, I don't understand Japanese so my re-interpretations might not be totally accurate but I try to match what's happening in the scene. Anyway, enjoy and let me know what you think​

 

Attachments

edited: On the latest version, the standalone .7z file comes with a One Click Transcribe.bat file. Now you can drag and drop file or files or folder onto this shortcut to do automatic one click transcriptions. So you can skip steps 3,4,5 below.

If you're on Windows, here is a standalone version:
  1. Go to Faster Whisper XXL Releases and download the latest Windows version.
  2. Extract the .7z file and navigate to the folder: Faster-Whisper-XXL.
  3. Place your movie file inside this folder.
  4. In the folder, type cmd in the address bar and press Enter.
  5. In the command prompt, type the following command:

    faster-whisper-xxl.exe "C:\yourmediafileslocation\abc-123.mp4" --language ja --model tiny --task translate --output_format srt​
6. This will start the process using your CPU.​

For NVIDIA GPU Acceleration:​

  1. Download the required CUDA files from CUDA Dependencies for Faster Whisper and extract all contents into the same Faster-Whisper-XXL folder.
    The files are:
    • cublas64_11
    • cublasLt64_11
    • cudnn_cnn_infer64_8
    • cudnn_ops_infer64_8
    • zlibwapi
  2. Run the same command as above and the process will start with CUDA, If it doesn't, add --device cuda to the command:

    faster-whisper-xxl.exe "C:\yourmediafileslocation\abc-123.mp4" --language ja --model tiny --task translate --output_format srt --device cuda

Additional Information:​

For more details about arguments and options, type the following in the command prompt:

Faster-Whisper-XXL -h​
This should get you started. For more accurate results, you might want to try the medium or large model with a capable GPU.
Thanks, it works nicely on my i7 with 4060t ti GPU. I ended going with large-V2. Takes anywhere bt 4 to 8 mins, Here's any example:

1733455132394.png
 
I have noticed that FasterWhisperXXL seems to produce more "hallucinations" than WhisperColab. I think I remember that WhisperColab had an update to try and limit hallucinations but I couldn't find the reference.
 
  • Like
Reactions: ericf
The colab uses a separate VAD system to handle the speech detection better than whisper alone and it also splits the audio in many small parts so that helps with hallucination.
 
Most if not all linked in this thread do, that's why all the ones linked in my tutorial(see signature) are called something like whisper with silero VAD.
 
  • Like
Reactions: Imscully
I have a good GPU (6 gb vram) but im using whisper on my computer using cmd prompt.
e.g. whisper --model large --language ja --task translate "C:\Users\user\Downloads\Video\1ha.mp3"

I want to use the VAD system instead, do you have a script or a tutortial on that?
 
Making a tutorial for that is in the todo list, but is has been for months and unlikely I'll have time to do it any time soon.

The colab has the code(commands) to install it on linux so you can look at that to see what you need to do/install, but might not be easy depending on how much technical knowledge you have.
 
I have a good GPU (6 gb vram) but im using whisper on my computer using cmd prompt.
e.g. whisper --model large --language ja --task translate "C:\Users\user\Downloads\Video\1ha.mp3"

I want to use the VAD system instead, do you have a script or a tutortial on that?
For Faster Whisper XXL, you can specify the VAD method to use with the --vad_method option.
 

Attachments

  • vad.jpg
    vad.jpg
    46.5 KB · Views: 20
The difference there compared to the colab is that you're doing the full movie in 1 shot. The longer the audio sample, the more whisper tends to hallucinate.

The silero VAD should detect speech better than whisper alone, but what helps with hallucination is splitting the audio.

I haven't tested this much myself, but it is what I hear from people.
 
  • Like
Reactions: ericf
Here's the colab result for the same portion you screenshotted with also the large-v2 model, certainly not perfect, lol.

Whisper_colab_aldn-399.jpg

I've always found whisper too random to compare, you can run the same audio with the same settings and get a different result.
 
  • Like
Reactions: mycl500
Here's the result from whisperjav 0.7

View attachment 3592576
This is great. It's a bit off in the highlighted text. To be fair, this part is always problematic. It's slightly better on XXL, but even there, it struggles with getting "he/she" correct. I tested WhisperJav as it is, without any additional settings or using DeepL. Do you mind sharing your settings? Thanks for the site!
 

Attachments

  • wj.jpg
    wj.jpg
    187.6 KB · Views: 19
Is whisperjav something you can run on your own computer or is is one of those online tools?
 
Seems soo much more accurate. I have a favour to ask, is there anyway you can teach me to set it up or the commands?
These are two vastly different whisper implementations. Whisperjav 0.7 has a lot more clean-up built in, which makes the immediate results seem better but the overall precision is lower due to the use of faster-whisper. Yes, it is faster and more cleaned right out of the colab but quality wise, the slower and slightly older WhisperPro is more precise. In the end, it's just your preference.