lmao def needs some tweaking
View attachment 549503
...
If that's taken from the tool in the opening post and it's set up properly, it definitely needs a lot more work or even a complete overhaul.
It looks like it struggles with capturing hiragana even though the text in that image is clear.
Somehow it seems to think くらす ('
to live' — as in '
living alone') means ころす ('
to kill') even though the characters look nothing alike. It also gets the name of the speaker wrong.
"
I'm finally going to kill you by myself." is borderline word soup; the words are English but the sentence is nonsense. Can you think of a single instance where you would say that exact sentence in English?
Another problem is that the text in the bubble is one sentence, but the translation turns it into two and just entirely fabricated the "
You did it, didn't you?" part because it interpreted the したんだな at the end as a separate sentence.
The two sentences added together also don't have any coherence, which puts it below even basic line-by-line translations done by human translators—which is one of the lowest forms of translation—as with those, you at least can tell what the original text meant.
Furthermore, the original line is a
statement aimed at the reader, but the translation turns it into
an action by the speaker, and the second sentence is even a question, and neither sentence even remotely conveys the meaning of the original text. It pretty much fails on every single level that it could have failed.
---
It's interesting that it's said to be aimed at retro games, which tend to have more kana than kanji, as that's one of the things it appears to fail at. Having it handle some normal text including kanji instead of the all-kana text like the above screenshots might produce more accurate results.
But from the screenshots in this thread, the translation ability seems very crude at this point, and it's definitely below something like Google Translate or DeepL, so you'd have to wait and see whether it improves in the future.
I'd say tools like that work best for single words like you'd see in command menus. For actual dialogue, the output tends to be subpar if you care about quality and accuracy. And at the level of those screenshots, it's worse than having no translation because almost everything is wrong and it would just confuse you.