Screen Translator

Features 🎯 In one line Game / manga / visual novel on screen → tap the floating ball → the translation appears over the original when recognition and translation finish, with optional read-aloud controls.

🖱️ How to trigger Tap the floating ball: defaults to translating the whole screen; switchable to Word-pick mode — one tap lets you draw a rectangle around just the word or phrase you want translated (more precise) Long-press the floating ball: opens an arc menu with common actions one finger can reach: loop / re-pick region / switch between "Translate full screen" and "Translate a word" / quick source-target language switch / preset switch / settings / back to main. Order and page size are configurable in settings; extra buttons paginate with a "Next page" button automatically Loop mode: choose a fixed interval or wait until dialogue text stabilizes; exact duplicate frames / translations are skipped, and detection can prioritize lower-screen dialogue regions for hands-free story reading Volume two-key: hold Vol+ and Vol− together for 0.3 s to trigger; your hands stay on the game (requires enabling the bundled accessibility service) From any other app: long-press to select some text in any app → tap "Screen Translator" in the system menu → translation card pops up immediately, no need to switch back to this app first 🔍 OCR (read the text on screen) On-device: ML Kit (CJK + Latin) / PaddleOCR / manga OCR — your screenshots never leave the phone; works offline; PaddleOCR supports multiple v5/v6 tiers Local HTTP OCR: connect to Umi-OCR / LunaTranslator services on your LAN or PC Cloud: PP-OCRv6 Online (PaddleOCR AI Studio) / Baidu / Tencent / Youdao — fall back to these when on-device OCR misreads; PP-OCRv6 Online requires an AI Studio Access Token When you switch source language, the app checks whether the current OCR engine can read it; if not, it suggests a better one Optional orientation detection can route horizontal / vertical / rotated text to a better OCR path, or you can lock the direction manually PaddleOCR, manga OCR, and orientation models select 1 / 2 / 4 / 6 CPU inference threads from the device's available cores, capped at 6 🌐 Translation engines LLMs: DeepSeek / ChatGPT / Claude / Zhipu / self-hosted services using OpenAI- or Anthropic-compatible connections, with results shown as they arrive On-device translation: Google ML Kit works offline after downloading its language packs; Sakura is tuned for Japanese manga and visual novels translated into Simplified Chinese, while Hy-MT2 supports more languages Consistent terminology: save names, places, organizations, and domain terms globally or per app so LLM and on-device model translations keep preferred wording across scenes DeepL: free / Pro plan auto-detected, supports self-hosted deeplx (open-source proxy, run on your own server, no key needed); pick between Official / deeplx / Auto-fallback Youdao Pic-Trans: skips the OCR step entirely — send the screenshot, get back boxes with translations. Great for comics. Google: no key, free (proxy required inside mainland China) Online engines have a Test connection action; DeepL can also show the current quota 🔊 Read translations aloud (TTS) Phone system voices: no account is needed. Use voices already available on the phone and adjust their speed and pitch More voice choices: use your own compatible voice service, or an online service such as Volcengine, MiniMax, or MiMo; online services may charge What it can read: source text, translations, word cards, and dictionary details. You can also select part of a horizontal passage and read only that selection Playback controls: tap the same passage again to pause or resume; tapping another passage switches to it, and failures are explained on screen. Some phone TTS engines do not report the current reading position, so resuming with those voices starts the passage again TTS is off by default and does not affect recognition or translation when unused 🔤 Word lookup (one tap = mini dictionary) Not just full-screen translation. Draw a rectangle around a single word or phrase and a card pops up with:

Translation — every engine returns this Pronunciation / part of speech / multiple definitions / example sentence pairs / difficulty notes — only when an LLM engine is selected; rare words, specialist terms, abbreviations, cultural references, and confusing usages are explained separately (Baidu / Tencent / DeepL etc. only return a plain translation) A "Copy source" and "Copy translation" button on the card; long-press any text inside to copy too In landscape, the close bar stays pinned at the top and both copy actions stay pinned at the bottom; only the content scrolls, clear of system bars and display cutouts Great for game / manga / VN vocabulary you don't recognize — one second per word, faster than switching to a dictionary app 🎨 How the overlay looks Two render modes: glued to each source box (BLOCKS), or packed into a draggable / resizable floating window — perfect for games with on-screen joysticks and buttons; you can lock the window to prevent accidental touches Adapt to screen: when translations are shown over the original screen, it chooses colors, background, and text size for each text area and places the translation back over the source; turning it off restores the previous display settings Unified layout and reading order: translated text follows OCR-detected horizontal / vertical layout and reading order by default; when disabled, both the layout and LTR / RTL direction must be selected manually, and OCR sorting uses the same resolved direction as rendering Selectable translation blocks: long-press a block for system text selection, or choose tap-to-open panel mode to select a range or copy the full source / translation 5 color themes + visual color picker: choose background, text, border color, and opacity directly instead of typing ARGB; the picker scrolls on short landscape screens Full text styling: bold, italic, underline, letter spacing, line spacing, alignment, outline, and shadow can be tuned independently, with a live preview Comic / subtitle optimizations: sentences split across multiple OCR boxes get merged before translating; vertical Japanese is read right-to-left; tiny ruby-text columns (furigana) next to kanji are filtered out so you don't get duplicate translations; vertical scenes can use vertical translation layout Marquee for long lines: in single-line mode, long translations scroll horizontally instead of being truncated with "…" Custom translation font: import .ttf files for translated text only; the app UI keeps using the system font, and imported fonts stay available as selectable chips 🛠️ Small conveniences English / 简体中文 UI + Light / Dark / Follow-system theme, applies instantly In-settings search: search in either language; new color, font, backup, import, and export settings are indexed too System presets: built-in bundles such as "offline Japanese manga OCR → Simplified Chinese" set up recognition, translation, and display in one step; the manga preset uses Adapt to screen by default, and lists any models you still need to download Background model downloads: continue after leaving Settings; both the app and system notification show the current model and progress, with cancel, retry, and resume support Per-app translation context: detect the foreground app and apply its terminology; package names always stay on-device, and only the display name is sent when you explicitly enable that option for a context-aware translator Portable settings bundle: export / import non-sensitive settings, custom presets, and font files in one step; API keys and tokens remain on the current device Translation cache: during the current process, identical text with the same translation configuration reuses its result for cloud engines and on-device LLMs Auto-saved crash reports: Java crashes, native crashes, ANRs, low-memory exits, and related failures are surfaced on the next launch as an exportable sanitized report Baseline TalkBack support: the floating ball, arc menu, language / preset switchers, region picker, word card, and color picker expose labels, state, and action hints Encrypted sensitive settings: API keys, prompts, mirror URLs and presets are migrated into encrypted local storage to reduce plaintext leftovers Update prompt: checks for new versions when you open the home screen (at most once a day); offers a direct link if GitHub is unreachable Chinese-ROM background guides: shortcuts to auto-start / battery whitelist settings for Xiaomi / OPPO / VIVO / Huawei / Samsung devices that aggressively kill background services Privacy default: cleartext HTTP is only allowed inside your LAN by default; public hosts must use HTTPS