AI translations going mainstream, can we have a list of text dumping and patching tools ?

[GengarsGhostlyGiggles]

AI translations are going mainstream apparently.

https://nichegamer.com/indie-srpg-little-hearts-gets-english-localization-thanks-to-ai-translation/
https://nichegamer.com/the-ancient-...imultaneously-in-english-with-ai-translation/

Between DeepL, ChatGPT (api), AWS translate, AWS Q Translate, Google Translate, Google Bard Translate, it seems the new neural network based models are extremely good now.
Facebook even released a next gen model https://about.fb.com/news/2023/08/seamlessm4t-ai-translation-model/ - SeamlessM4T, that you can actually just download from huggingface (https://huggingface.co/docs/transformers/main/model_doc/seamless_m4t) and RUN ON YOUR HARDWARE FOR CHEAP!.

On twitter, some people have been saying they even have ways to have AI do the "photoshopping" on textures.

I was wondering if people could give me some recommendations for tutorials/software on dumping text, textures and then patching the translated stuff back in.

Post automatically merged: Dec 24, 2023

Also does anyone know how people are using ai "photoshopping", is that just the official adobe photoshop ai ?

 

[masagrator]

Those AI translations are proofread by human translators. So no, they are not getting good enough to be used blindly. Especially for any other language than English.

And you need to figure out how to configure AI before start of translation so it knows the context of every text, otherwise it likes to output garbage.

 

[GengarsGhostlyGiggles]

Those AI translations are proofread by human translators. So no, they are not getting good enough to be used blindly. Especially for any other language than English.

And you need to figure out how to configure AI before start of translation so it knows the context of every text, otherwise it likes to output garbage.

yea I'm working on a process that can compare a single line across all AI's at once, then streamline the human review process kinda like this tool https://prodi.gy/.

I know how to code roughtly, so it's not hard at all, but i'm thoroughly unfamiliar with how text/textures get dumped and patched.

I've been testing against some visual novel csv text dumps, and have really good results.

 

[masagrator]

I will give you a chance to convince me.
Choose some VNs released on Switch that have good scores for plot (they don't need to be kamige ), have voice acting and were not translated for any platform + it doesn't have any active translation going on. I will choose 1 that won't conflict with other projects I'm working on. I will try to make tools for it to allow you translating text. If translation quality will pass my shitty knowledge of Japanese and won't occure to me easily that something is wrong, I can help with other VNs.

To make it easier, list excluding released non-machine English translations:
https://vndb.org/v?q=&ch=&f=032hen3gja4oswi&s=24w

It still lists ongoing and machine translations, so you must check each entry inside.

 

[GengarsGhostlyGiggles]

I will give you a chance to convince me.
Choose some VNs released on Switch that have good scores for plot (they don't need to be kamige ), have voice acting and were not translated for any platform + it doesn't have any active translation going on. I will choose 1 that won't conflict with other projects I'm working on. I will try to make tools for it to allow you translating text. If translation quality will pass my shitty knowledge of Japanese and won't occure to me easily that something is wrong, I can help with other VNs.

Hmmm okay sounds good.
Does this workflow make sense to you ?
1. generate dumps of the text
2. create application that takes the dumps
3. breaks them up into paragraphs
4. run the paragraphs + any additional context thru 4-5 translation AI's
5. store results in a sqlite database
6. Run some statistics to assemble a good idea of how similar or dissimilar the various ai's are for the same paragraph to help automate the quality checking or control.
7. Allow a human to then grade the translations or select the best via a GUI tool like web ui or electron gui. or store results in the sql lite database.
8. Take the best graded translations for a given game and re assemble the text into a dump for patching ?

I'm not sure how people how are doing ai photoshopping just yet.

 

[masagrator]

Does this workflow make sense to you ?

Well, I'm not a translator. I'm only a person that likes to find patterns in hex editors with help of debugger and disassembler + write programs to make them editable. I am interested only in final result.