Do all 23 languages get the same features?

No. Translation works between any of them, but text-to-speech covers 6 languages (English, Japanese, Spanish, French, Italian, Portuguese), and difficulty-tag vocabulary data exists only for Japanese, English, and Chinese. The Language Support page lists what each language currently gets.

How can the source and target languages be combined?

Whichever of the 23 you pick as the source, you can read it in any of the remaining 22 — 506 combinations. Play a Japanese game and read it in English, play an English game and read it in Spanish, and so on.

Translation quality is poor for my language. What should I do?

First switch the translation model in Settings and run the same scene again. Local models each have languages they handle well and languages they don't, so results vary by language pair. If that doesn't help, send us the game, the kind of screen, the translation you actually got, and the model you were using, through Contact us in the app. Testing every combination ourselves isn't realistic, so those reports are the main thing that drives fixes.

Playto now supports 23 languages

Playto now supports 23 languages. This update adds Dutch, Swedish, and Traditional Chinese, and each of them works as a source language or as the language you read in — 506 combinations in total.

The bigger part of this update probably isn’t the three new languages, though. It’s the fix that shipped alongside them.

Words with accented characters were vanishing

In Swedish or Finnish games, words containing å or ä could be on screen and still never reach the translation step. It wasn’t that the translation was poor — the word never arrived.

There were two causes. The first was a filter that drops low-confidence results from text recognition: it was meant to remove punctuation and noise, and it was taking real words with it. Every non-CJK language was affected, all 19 of them. In our measurements, 120 of the 134 dropped items were real words.

The second cause was on the image side. Before a captured region reached the OCR engine, one path still converted it to grayscale and stretched its contrast, and that flattened the accent marks. The code said the image was being passed through untouched; it wasn’t. On the same screenshot, 16 broken spots before the fix are down to one or two characters now. Not zero, so we’re still chasing the rest.

For a long stretch, European languages must have felt vaguely off in a way that was hard to point at. It took us too long to find this.

Languages don’t all get the same features

To be straightforward about it: the 23 are not equal.

Text-to-speech covers six languages — English, Japanese, Spanish, French, Italian, and Portuguese. The lightweight NMT model, the one that runs on modest hardware, covers a fixed set of source languages, and Swedish, Romanian, Danish, Norwegian, Finnish, and Hungarian aren’t among them, so those always go through AI translation instead. Difficulty-tag vocabulary data exists for Japanese, English, and Chinese only.

We’ve put the per-language picture on the Language Support page. If you want to know what your language gets today, that’s the fastest place to look.

If a translation reads badly, try a different model

The other thing that varies by language pair is the model itself. The small models that run locally each have languages they handle well and languages they don’t, so the same settings can produce solid translations in one pair and mush in another.

Playto preselects a recommended model based on the language you choose, but that pick isn’t always the best one for your pair. When a translation reads badly, switch models in Settings and run the same scene again. That alone often changes the result. The models we bundle, and how to choose between them, are covered in Models.

If that doesn’t help, tell us through Contact us in the app. The game, which screen it was, the translation you actually got, and the model you were using are what let us reproduce it here. That’s how the recognition bug above surfaced.

What’s next

Adding more languages has come off the roadmap. We think evening out the quality across the languages we already cover matters more than the number 23. What we are prioritising instead is public on the roadmap.