• Hey Guest. Check out your NeoGAF Wrapped 2025 results here!

Interpreter is an open-source tool that uses AI to translate Japanese retro games on the fly

IbizaPocholo

NeoGAFs Kent Brockman

Developer Bertrand Quenin recently released an open-source project called "Interpreter" that aims to provide real-time translation for Japanese retro games. The tool can capture Japanese text from virtually any on-screen window, perform live optical character recognition, and render an English translation directly over the game as it runs.

Interpreter operates entirely offline, handling OCR and translation locally without relying on cloud services or paid APIs. The tool is free to use and inherently private, as captured text never leaves the user's machine. Translations are displayed using two overlay modes: a banner mode that functions like subtitles, and an in-place mode that renders English text directly over the original Japanese.

Two other open-source projects lie at the heart of Interpreter's OCR and translation capabilities. MeikiOCR accurately extracts Japanese text from video games, while the large language model Sugoi V4 converts Japanese ideograms into English.

Interpreter runs on all major PC operating systems, including Windows, Linux, and macOS. On Windows, users install the Python-based project by running a PowerShell script. The developers caution that initial OCR accuracy may be inconsistent and that first-time launches can be slow. Adjusting a few settings in the graphical interface and caching the OCR and language models locally typically resolves both issues.

Interpreter is an intriguing project, but it still faces compatibility and reliability issues. Some users suggest it could be far more helpful and efficient if integrated directly into emulators as a plugin. For now, the tool appears to work on the PS3 emulator RPCS3, but only in windowed mode.


Thanks, mods, for the title update :messenger_clapping:
 
Last edited:
That's pretty interesting implementation and does make sense. I have used text hooks to get data out for sentence mining but that usually made calls to a dictionary resource.

A Language model does seem ideal for this. I do wonder about accuracy especially with a small local model.
 
You can't match every single word meaning a %100 adding the dialect on top of all that and you end up with a complete messy situation no one could ever solve, but technically speaking the native speaker must do all the hard work to get it closer to perfection.
 
Last edited:
You can't match every single word meaning a %100 adding the dialect on top of all that and you end up with a complete messy situation no one could ever solve, but technically speaking the native speaker must do all the hard work to get it closer to perfection.
I imagine they are shooting for standard "Tokyo" Japanese. Maybe they trained on Kansai-Ben, but I kind of doubt as that's a lot of work.

DeepL is relatively (key word here 😀) now days with Japanese translation and there are specialized models (like this one) that are further trained.
 
Last edited:
Top Bottom