Droid Sentence Miner

Android native sentence miner: popup OCR + Yomitan dictionaries + Anki

A native Android sentence miner for Japanese immersion: OCR text from any app with a floating popup, look words up with Yomitan dictionaries, and mine them into Anki — plus the game's voice line as sentence audio.

Inspired by GameSentenceMiner, rebuilt for Android with Yomitan dictionary compatibility.

Book reader: tap a word on a manga page

Features Auto OCR mode — draw boxes over the game's text areas once (dialogue box, name box, subtitles; kept separately for portrait and landscape) and keep playing. The areas are read every ~400 ms; each new line is logged with the moment it appeared, the moment it went away and a screenshot. Repeated reads must agree before a line counts, and the stored text is the consensus of all reads of that line, so fades, typewriter text and OCR noise over moving video don't create junk; background signage is filtered out. Visual novel text that types itself out a few characters at a time is followed to the end of the line (each frame is compared with the last one OCR'd, cell by cell), so lines aren't cut mid-sentence. Screen dictionary (tap the real text, in any app) — while auto OCR runs, each read is laid as an invisible text layer exactly over the subtitles/dialogue on screen (meikiocr's and ML Kit's per-character positions). Tap a character in a YouTube hardsub or a game's text box and the word covering it lights up with its definition right next to it; hold and drag to select characters. The layer is one invisible window per line, fitted to it, so touches elsewhere (a game's buttons, the empty part of a text box) still reach the app. Videos are fast, so a lookup pauses playback (media key: YouTube and most players) and resumes it when the popup closes; if you pause yourself, the layer stays and you can tap word after word. Mining links the log line, so the card gets the sentence, screenshot and voice clip. Already in Anki — like Yomitan, the popup checks your note type for the word: if it already has a note the button shows ✓ In Anki instead of making a duplicate. A card keeps being added if you close the popup right after tapping Anki. Tablets and landscape — tabs move to a side rail and screens sit in a centered column. Live text window — a floating feed of the lines over the game, updated in real time. Tap a word in it and the definition pops up right next to it (same popup as the on-screen text layer; mining uses that line's audio and screenshot); hold a line to open it in the full panel. Scroll back through the history (new lines only pull it down while you're at the bottom; Latest jumps back), drag the corner grip to resize, shrink it to a one-line ticker (⌃⌄), or open it fullscreen with big text to read back (auto OCR pauses there, since it would cover the game). Size and position are remembered per orientation (a big portrait window would cover a landscape video), and so is the mode; fullscreen always reopens as the normal window. App focus — share "A single app" (Android 14+) and pick the game: only it is read, and auto OCR pauses while you are in other apps or in Droid Sentence Miner itself. Book mode (paper books, manga) — the Book tab takes a photo with Samsung Camera on Galaxy phones (any camera app elsewhere) or picks one from the gallery, reads the page with ML Kit — vertical (tategaki, columns right to left) and horizontal print — and outlines each paragraph on the photo. Tap a paragraph, tap a word, and mine it: the card gets the word's sentence (cut at 。！？, keeping closing quotes) and the paragraph cropped from the photo as its picture. Hold and drag over the text to force your own selection (hold alone = one character, e.g. a single kanji). The definition pops up right at the word (below it, or above when there is no room), instantly and without a scrim, so tapping the next word just moves it. Tap Japanese inside a definition to look that up too (← goes back). Ruby lines, page numbers and sentences split across blocks are handled; ✎ fixes a misread before mining. Overlay dictionary mode (read on the photo) — Mokuro-style: after a photo the page opens full screen with the OCR text laid invisibly over the printed characters (ML Kit's per-character boxes). Tap any character on the photo, in a speech bubble or a column of print, and the word starting there lights up with its definition popping up beside it; hold and drag to select characters yourself; pinch to zoom, drag to pan. One add button per reading, like Yomitan — when a lookup finds several headwords (呼び and 呼ぶ, 取り and 取る, different readings of one kanji), the popup lists each with its own + Anki button and in-Anki state; the card gets only that headword's reading, glossary, pitch and frequency. The header button adds the first one. Mining feedback — the popup's Anki button spins while the card is built, turns green (Added) when it's in, and a toast confirms it (or says why not) even if the popup was closed. Texts page (GSM's texthooker) — every recognized line, newest first. Tap a line, tap a word, and mine it: the card gets that line's sentence, its screenshot and its voice clip. Replay buffer sentence audio — the last 5 minutes (configurable 1–10) of game audio are kept in memory. Any line still inside the buffer can be mined with audio cut from when it appeared to when it was replaced (± configurable offsets). Preview or fine-tune it in the waveform editor before mining; optional (experimental) silence trimming is in Settings. Floating popup overlay — a draggable bubble that stays on screen. Tap it, drag a box over a word and its definition pops up right next to it (Quick lookup, the default); other words in the box can be tapped too. Long-press opens the menu, where Bubble tap switches to Scan region (the region editor with Scan / Full screen / Live and the results panel). The notification stays available as a shortcut. Japanese OCR trained on games — on-device meikiocr (ONNX Runtime), a detector + recognizer trained on Japanese video-game text: 3–19× fewer character errors than ML Kit on stylized, outlined game/subtitle fonts (small kana, dakuten). One-time 31 MB model download; ML Kit (with a GameSentenceMiner-style two-pass pipeline) is built in as the fallback/alternative engine. Yomitan dictionaries — import any Yomitan/Yomichan format 3 .zip (Jitendex, JMdict, KANJIDIC, frequency dictionaries) or install Jitendex/JMdict/KANJIDIC with one tap. Term, kanji, tag, term-meta and kanji-meta banks are all parsed, including structured-content glossaries and dictionary images. Updatable dictionaries (Jitendex, or any Yomitan dictionary with an indexUrl in index.json) are checked for newer revisions automatically and can be updated in place with one tap, preserving priority and enabled state. Deinflection & word scanning — a full Japanese deconjugator drives longest-match scanning of scanned text and tap-to-look-up, with the same rules Yomitan uses. Anki mining — two backends: AnkiDroid API (no network needed; cards, images and audio via the official API), and AnkiConnect (desktop Anki with the AnkiConnect add-on; card + media upload over HTTP). Cards include Expression, Reading, Meaning (rendered glossary), Sentence, Image (captured screenshot), Audio (sentence voice line), Word audio and Source. Existing note types are filled by field name the way Yomitan/GSM fill them — e.g. Anime Cards: Word, Reading, Glossary, Context Sentence (Sentence stays empty so it stays a word card), Word Audio, Audio, Picture. Settings → Anki → Fields shows what goes into each field of the chosen note type and lets you pick the content per field (Word, Reading, Glossary, Sentence, Screenshot, Sentence/Word audio, Pitch graph, Source or nothing); the mapping is saved per note type, Reset to automatic drops it. Word audio like Yomitan — the word's pronunciation is fetched from the same sources Yomitan uses for Japanese, in the same order: JapanesePod101 (placeholder clip rejected by hash), JapanesePod101 dictionary search, Jisho; plus an optional custom URL ({term}/{reading}, e.g. a local audio server) tried first. Option to put only the first dictionary on the card (Yomitan's {glossary-first}). Pitch accent graph — the same SVG Yomitan's {pitch-accent-graphs} writes (byte-identical to the user's existing Yomitan cards), filled into a pitch/graph field such as Anime Cards' Graph. Uses any installed Yomitan pitch dictionary, or the free Kanjium pitch accents (124k words, one-click install from Settings → Dictionaries or the Home setup card). Sentence audio (Android 10+) — captures playback audio through the existing MediaProjection session (the Android equivalent of GSM's loopback capture) into the replay buffer, with a waveform audio editor to trim the exact line before mining. Every clip is normalized to the same loudness (on by default, Settings → Sentence audio). Microphone fallback — some apps block audio capture and Android hands over pure silence (DroidDeck, which runs Steam games, targets Android 9; Netflix). Turn on Microphone fallback (Settings or the bubble menu; off until you allow it) and DSM records the microphone whenever device audio is silent while something is playing, and switches back as soon as device audio returns. The mic goes through an FX chain built for a speaker heard through a microphone (each stage on by default, can be turned off): Android's voice noise suppression, a speaker correction EQ — a 120 Hz rumble filter plus a 19-band curve measured on your device by Calibrate (6 s of pink noise played and recorded back), which restores the missing low mids and tames the harsh 4–8 kHz resonance of tablet/phone speakers — then a compressor that squeezes loud peaks so normalization can lift thin, quiet dialogue. Works with the speaker on; picks up room noise; records nothing with headphones. Mining history — recent cards with a status showing whether they made it into Anki. Text selection lookup — other apps can send selected text via ACTION_PROCESS_TEXT, including Samsung's text selection on screenshots/images and Circle to Search. A selected word is matched to the newest logged line containing it, so the card still gets the full sentence, screenshot and audio; a selected sentence is logged as a new line. Handy for text auto OCR can't read (vertical or handwritten-style), using Google's cloud OCR by hand.

License

GPL-3.0 license