Best program to use for MTL?

[TheSoaringDingo]

I've been hearing a lot about AI and the fact that MTL are getting significantly better to the point where Crunchyroll is actively using it. I was wondering what MTL program people would recommend for me if I wanna try to patch a VN. 

[Entai2965]

Most AI API providers assume you can write basic code yourself. Even though about a year and a half has passed since AI really took off, this is still very early in terms of software development since software development takes time, especially for quality software. There are not a lot of good options right now, and almost all of them require tinkering and being very familiar with the particulars of both the model you are using and the API you are working with. A short 100 line python program could do it, but you would have to plug in your data manually and format it or whatever.

Some examples are DazedMTLTool, Translator++, SLR Translator, and a handful of other projects on github in various states of development. Dazed is ChatGPT only. Translator++ is DRM'd paid software that requires constant internet connection to validate your patreon account, the AI addons cannot be used otherwise. SLR Translator is not yet released but should have some AI addons shortly after it is released. Shortly means like within a few months.

DeepL has also started to incorporate AI into their API as well, so you can just submit a prompt that has extra information for DeepL to consider when generating their output. They are not currently charging for this but like... they are going to very soon, so use it while you can. DeepL is free for the first 500k characters/month. All other APIs are paid/rate limited. If you want to do unlimited, then you basically have to run AI yourself locally using KoboldCpp or similar programs and write your own software to interface with it. That said, IDK of any software that has working DeepL API + prompt translations working, so you are back to writing it yourself territory there.

If you just want basic translations, NMTs are much easier to use. Personally use my repackage for the Sugoi v4 NMT model. It has an easy to use Web UI written by one of the server authors, so I just dump the raw text in there and get it back right away. But NMT models, like Sugoi v4, are not AI obviously, they are the precursor to contemporary AI. They are basically free though 500+ lines in ~30s to calculate w/CUDA, and they provide a good basis to create edited MTL patches as long as you are willing to take a bit of time to format the output yourself.

Edited MTL is better quality than AI generated TLs because it means a human looked at the output, so do not look down on NMTs too much. Just because AIs are the superior successor technology does not mean they produce superior output without A LOT of fiddling and certainly they do not compare well against edited NMT translations at all.

In terms of quality: Human translator > Edited MTL (AI or NMT) > MTL (AI) > MTL (NMT) > MTL (Dictionary)
In terms of cost and ease of use: MTL (Dictionary) > MTL (NMT) > MTL (AI) > Edited MTL (AI/NMT) > Human translator

Edited May 10, 2024 by Entai2965

[TheSoaringDingo]

5 hours ago, Entai2965 said:

Most AI API providers assume you can write basic code yourself. Even though about a year and a half has passed since AI really took off, this is still very early in terms of software development since software development takes time, especially for quality software. There are not a lot of good options right now, and almost all of them require tinkering and being very familiar with the particulars of both the model you are using and the API you are working with. A short 100 line python program could do it, but you would have to plug in your data manually and format it or whatever.

Some examples are DazedMTLTool, Translator++, SLR Translator, and a handful of other projects on github in various states of development. Dazed is ChatGPT only. Translator++ is DRM'd paid software that requires constant internet connection to validate your patreon account, the AI addons cannot be used otherwise. SLR Translator is not yet released but should have some AI addons shortly after it is released. Shortly means like within a few months.

DeepL has also started to incorporate AI into their API as well, so you can just submit a prompt that has extra information for DeepL to consider when generating their output. They are not currently charging for this but like... they are going to very soon, so use it while you can. DeepL is free for the first 500k characters/month. All other APIs are paid/rate limited. If you want to do unlimited, then you basically have to run AI yourself locally using KoboldCpp or similar programs and write your own software to interface with it. That said, IDK of any software that has working DeepL API + prompt translations working, so you are back to writing it yourself territory there.

If you just want basic translations, NMTs are much easier to use. Personally use my repackage for the Sugoi v4 NMT model. It has an easy to use Web UI written by one of the server authors, so I just dump the raw text in there and get it back right away. But NMT models, like Sugoi v4, are not AI obviously, they are the precursor to contemporary AI. They are basically free though 500+ lines in ~30s to calculate w/CUDA, and they provide a good basis to create edited MTL patches as long as you are willing to take a bit of time to format the output yourself.

Edited MTL is better quality than AI generated TLs because it means a human looked at the output, so do not look down on NMTs too much. Just because AIs are the superior successor technology does not mean they produce superior output without A LOT of fiddling and certainly they do not compare well against edited NMT translations at all.

In terms of quality: Human translator > Edited MTL (AI or NMT) > MTL (AI) > MTL (NMT) > MTL (Dictionary)
In terms of cost and ease of use: MTL (Dictionary) > MTL (NMT) > MTL (AI) > Edited MTL (AI/NMT) > Human translator

wow that's a lot of info. Thank you for helping me out! I will look into the methods you recommended