playtranslate

PlayTranslate is an Android application for real-time translation of video-game text. It is aimed at language learners and players who want to read games in languages they do not speak. The project is distributed as a sideloaded APK, with the README noting that Google Play Protect may need to be temporarily disabled on some devices, and that certain OEMs require enabling "Allow restricted settings" before accessibility permissions can be granted. Core capabilities described in the README: - One-tap capture and translation of the game screen, plus an auto mode that re-translates as dialogue changes. - A floating lens for word lookup, showing dictionary definitions on hover. - Offline operation: OCR and dictionary lookups work without a connection, and optional offline translation models can be downloaded. - Reading aids: furigana and pinyin hints displayed above characters. - Hotkey support for hold-to-preview translations or reading hints, intended for handhelds with dedicated buttons. - Dual-screen and split-screen support, including dual-display devices and Android split-screen with windowed games. - Capture regions to limit OCR to a dialogue box, subtitles or a custom area. - Text-to-speech playback with a configurable voice. - Anki export to AnkiDroid, including original text, translation, word list, target words, TTS, screenshot and optionally recorded game audio, with card-type presets. - Yomitan dictionary integration, including pitch accent, frequency chips, kanji enrichment and merged definitions, also used in Anki exports. - Camera translation: live reading of text in the world, or freezing a frame to tap and look up words. - Optional text history, off by default. Language coverage is stated as 26 game (source) languages read from the screen and 59 translation (target) languages, with tables listing codes and native names for each. Translation backends: by default the app uses Lingva with ML Kit as an offline fallback. Users can add API keys for DeepL, OpenAI, Gemini, DeepSeek, Mistral, Groq, OpenRouter, Claude, or any OpenAI-compatible custom endpoint, and reorder services to control priority. Optional downloadable offline models mentioned include Gemma 4 E2B, Hy-MT2 1.8B, Qwen 3.5 2B, retired Qwen 2.5 1.5B Instruct and Hunyuan-MT 1.5 1.8B packs, per-script PaddleOCR recognizers, the Meiki Japanese OCR model, MangaOCR for stylized and vertical Japanese text, and Firefox Translations (Bergamot) model pairs. The README credits a long list of libraries and data sources, including ML Kit, Sudachi, HanLP, KOMORAN, Snowball stemmers via Apache Lucene, Lingva, AnkiDroid, MNN, PaddleOCR, OpenCV, Silero VAD, OpenCC4j, slimt/Marian, OkHttp, kotlinx.serialization, Gson, CameraX and Material Components. It also documents work adapted from offline-translator, PyThaiNLP, docTR, EasyOCR and Yomitan, and linguistic data from JMdict, KANJIDIC2, JMnedict, CC-CEDICT, CFDICT, HanDeDict, Wiktionary via kaikki.org, Tatoeba, PanLex, wordfreq, Camel Morph MSA, Arramooz, Morfologik/PoliMorf, SudachiDict, Jiten and Wikimedia Commons. The project is licensed under GPL 3.0. Support is offered through a Discord server, with donations via Ko-fi.