Frequently Asked Questions

Playto is a language learning tool for PC gamers — not a translator. As you play, it reads on-screen text, shows what it means, and lets you save words to a personal dictionary for later review.

The more you play, the more your personal dictionary grows.

The overlay gives you meaning hints so you can keep playing. The real value is in building your own vocabulary over time and reviewing it with the built-in quiz and SRS system.

Honestly, quality varies by language pair:

• English ↔ Japanese (both directions) — our primary tested pairs. Most tuning, evaluation, and prompt work has been done here.
• English ↔ Chinese / Korean — well-tested via the CJK OCR path.
• English ↔ European languages (es/de/fr/it/pt) — functional but less tuned; casual-use quality, long sentences can be rough.
• Russian, Polish, Vietnamese — supported as language pairs but minimal testing. These need community help.

See "How can I help improve Playto?" below.

Three ways to contribute:

1. Report bad translations — which game, what text, what you expected. We use this to refine per-language prompts.
2. Contribute to game glossaries — character names, skill names, world-specific terms that AI models can't infer.
3. Test your language pair — especially if marked "less tested" above, your feedback is especially valuable.

We can't professionally test all 132 language pairs (12 × 11), so community feedback fills that gap. Join our Discord.

Playto combines several open-source components, each chosen for a specific job:

Text recognition (OCR)

• Windows Media OCR — system-provided engine, used by default on Windows

Translation & meaning hints (LLM)

• Qwen 3 series (Apache 2.0) — default
• Gemma 3 series (Gemma Terms of Use) — selectable alternative
• Swallow 8B (Llama Community License) — Japanese-specialized option
• Qwen 3 VL series (Apache 2.0) — used in Vision mode (direct image understanding)

Language analysis (tokenization / morphology)

• kuromoji.js (Apache 2.0) with IPADIC — Japanese morphological analysis
• Intl.Segmenter (ICU / Unicode License) — word segmentation for other CJK languages

Speech (TTS & pronunciation check)

• Kokoro TTS (Apache 2.0)
• espeak-ng (GPL-3.0) — IPA phonetic generation
• whisper.cpp (MIT) — pronunciation test (speech recognition)

Runtime

• llama.cpp (MIT) — local LLM inference
• ONNX Runtime (MIT) — Kokoro TTS and Vision model inference
• Tauri v2 (Apache 2.0 / MIT)
• Svelte (MIT)

Optional cloud mode — if your local GPU is tight on VRAM, you can choose Google Gemini or OpenAI from Settings. In that mode only the text to be translated is sent to that provider. See the Privacy Policy for details.

Model selection and components may evolve (additional language analysis libraries are planned as more languages are supported) — this list reflects the current shipping configuration.

OCR works best when the text is clear and high-contrast. Try these fixes:

• Use Borderless Windowed mode instead of Fullscreen. This gives Playto cleaner screen captures.
• Draw the capture region tightly around the text area — avoid including game graphics.
• If the font is stylized or decorative, try Image Recognition (Vision Mode) instead — it reads the screen as an image and handles stylized text much better.
• Make sure the game is running at your monitor's native resolution.

Honest answer: long-sentence translation is a work in progress.

Why it's hard: AI models translating game text don't know character personalities, speech tone, or storyline context. Even large models struggle with disconnected lines — a single sentence like "You're weaker than you think" can mean very different things depending on who's saying it.

What helps:

• Treat the overlay as a meaning hint, not a definitive translation — read the original text alongside it.
• Use the dictionary popup for word-level lookup — this is reliable.
• Add character/term glossaries for games with lots of proper nouns.
• If VRAM allows, try a larger model (e.g., 13B instead of 7B) for smoother long-form output.
• Cloud AI can be more fluent but doesn't solve the context problem.

This is our biggest ongoing improvement area. Community feedback helps us refine per-game and per-language prompts.

Translation speed depends on your AI model and hardware. Try these:

• Check GPU Layers in Settings > Model. Make sure the model is offloaded to your GPU. More GPU layers = faster inference.
• Try a smaller model (e.g., 7B instead of 13B parameters). Smaller models are faster with only a slight quality trade-off.
• Use Cloud AI if you have a weak GPU. Cloud APIs (OpenAI, Gemini, etc.) are fast and don't use local resources.
• Close other GPU-heavy applications. Your game and the AI model share VRAM.

Playto has a window detection feature that should prevent this. If it's happening:

• Make sure your game window is focused (active) when Playto captures.
• Check your Game Pack settings — the process name should match your game's executable.
• The overlay window itself is excluded from captures automatically.

Your game and the AI model share your GPU's VRAM. If you're running out:

• Use a smaller model. A 7B model needs ~4GB VRAM. A 3B model needs ~2GB.
• Reduce GPU layers. Partial offloading uses less VRAM (the rest runs on CPU, slower but works).
• Switch to Cloud AI. Cloud APIs use zero local VRAM.
• Lower your game's graphics settings to free up VRAM.

TTS availability depends on the language pair. Check:

• Make sure TTS is enabled in Settings > Audio.
• Confirm your target language has TTS support. Most major languages are supported.
• Check your system audio output — make sure it's not muted or routed to the wrong device.

Pronunciation uses a local Whisper model for speech recognition:

• The Whisper model downloads automatically on first use (~140MB). Make sure the download completed.
• Check that your microphone is connected and has permission in Windows settings.
• Pronunciation runs on CPU only (to avoid GPU contention with your game), so it may take a moment to process.

Translation processing runs locally on your PC — you don't need internet while playing.

• Local AI mode (default): OCR, translation, and all learning data stay on your machine.
• Cloud AI mode (opt-in): If you enable this, the text being translated is sent to your chosen provider (Gemini, OpenAI, etc.). Your dictionary, quiz history, and glossaries are never sent.

What we don't do:

• Don't collect your dictionary, quiz history, or screenshots.
• Don't send analytics by default.
• Don't require an account.

What we might do in the future (opt-in, with explicit consent):

• Game vocabulary difficulty aggregation (counts only, no individual word lists) for a public CEFR Games database.
• Community glossary sharing to help others learning the same game.

Any future data collection will be clearly disclosed, opt-in by default, and visible in Settings.

$9.99 on Steam, one-time purchase, with a 10% launch discount for wishlist users.

All learning features are included: OCR/Vision translation, cursor and fixed-region capture, dictionary, 8 quiz modes, SRS review, text-to-speech, glossaries, journal, achievements, and the upcoming mobile companion app. No subscription.

In development. The plan: a free mobile companion bundled with your Steam purchase, focused on reviewing words you saved during PC play.

Sync will be local-network only — no cloud accounts, no subscription, and your learning data never leaves your devices. Release target: later this year.

Any PC game that displays text. Playto reads your screen and works as an overlay — no modding, no SDK integration, no patching required.

Text-heavy genres work best: JRPGs, Visual Novels, Strategy games, Simulation, MMOs. Action-heavy games work too, but you'll have fewer words to save per session.

We haven't tested every game. Community-tested titles and glossaries are shared in the Discord #games channel.

Playto only reads screen captures — it doesn't inject into game memory or modify game files. However, some aggressive anti-cheat systems flag any kind of screen capture software.

Recommendation: check the game's anti-cheat documentation or test in a secondary account first if you're concerned. Community reports on specific games are welcome.

Post in the Discord #game-requests channel with:

• Game name
• Which store (Steam, Epic, GOG, etc.)
• Language pair you want

Popular requests get community glossaries faster. You can also help by sharing your own glossary entries once you start playing the game.