Droid Sentence Miner
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. Japanese is the focus; English and German are supported too, and other languages can
be added as modules that download nothing unless you study them.
Inspired by GameSentenceMiner, rebuilt for Android
with Yomitan dictionary compatibility.
Website: wilsonnet.github.io/droid-sentence-miner ·
User docs: Installation ·
Auto OCR ·
Troubleshooting ·
Changelog
Demo
| Visual novel (Nightshade on Steam via DroidDeck) | YouTube hardsubs |
|---|
| |
| Tap 手薄 in the game's text box → definition → Anki → ✓ In Anki | Tap 消防車 in the burned-in subtitle → Anki → ✓ In Anki |
Both cards got the sentence, a screenshot, word audio and the sentence's voice clip (click a
GIF for the MP4; recorded on a Xiaomi Pad 7).
What it does
- Auto OCR — box the game's dialogue or a video's subtitles once; every new line is logged
with its screenshot and voice clip, using OCR trained on game fonts
(meikiocr, ML Kit as fallback).
- Tap the real text — an invisible layer over the on-screen text lets you tap a word in any
app and get its definition right there. A floating bubble, a live text window and a Texts log
do the same.
- Yomitan dictionaries — Jitendex, JMdict, monolingual dictionaries, pitch and frequency
dictionaries, with images and dictionary CSS. Install them with one tap, from files or from
Google Drive; reorder by dragging.
- Anki cards like Yomitan/GSM make them — through AnkiDroid or AnkiConnect: sentence,
screenshot, sentence audio, word audio, pitch graph, mapped to your existing note type.
- Sentence audio — a replay buffer of the app's audio, a waveform editor, loudness
normalization, and a calibrated microphone fallback for apps that block capture.
- Books and manga — photograph a page and tap the printed text to look it up.
- Multilingual, Japanese first — pick the language you study on first launch: Japanese
(main focus), English or German (Wiktionary dictionaries, inflections, IPA, German articles on
cards, optional PaddleOCR pack); each language has its own Anki deck, note type and card
contents (dictionaries per language).
Only that language's dictionaries and models are downloaded. New languages are small modules
(how).
Every feature in detail: docs/features.md.
Get it
Android 8.0+ (sentence audio needs Android 10+), plus AnkiDroid
or desktop Anki with AnkiConnect. Signed APKs are on
GitHub Releases; newer releases
upgrade in place.
Quick start: Settings → Dictionaries → install Jitendex · Settings → OCR → download
meikiocr · Settings → Anki → pick a backend, deck and note type · start the bubble, open your
game, long-press the bubble → Start auto OCR. The full walkthrough is in the
installation guide
and docs/usage.md.
Documentation
Contributing
Everything builds from the command line — no Android Studio needed:
./gradlew testDebugUnitTest assembleDebug # JDK 21 + Android SDK, see docs/building.md
adb install -r app/build/outputs/apk/debug/app-debug.apk
./gradlew detekt detektDebug lintDebug testDebugUnitTest is the PR gate: every detekt/ktlint
rule, every Android Lint check, and architecture tests that keep language modules separate
(details). Debug builds include a localhost HTTP API for driving and checking the app over adb, so most
changes can be validated with curl on a real device. Before opening a PR, read
AGENTS.md: it lists the pitfalls already paid for, and every change updates the docs
in the same commit (checklist). The website lives in
site/ (Astro + Starlight).
Credits
This project stands on the shoulders of these projects — thank you:
- GameSentenceMiner (GPL-3.0) — the whole
concept: OCR on untranslated games, popup lookups, one-press Anki cards with audio, the two-pass
OCR pipeline and the sentence-audio editing workflow.
- Yomitan (GPL-3.0) — the dictionary format, deinflection
approach, structured-content glossaries and popup UX that this app implements for Android.
- Yomihon (Apache-2.0) — the Kotlin Japanese deinflector in
JapaneseDeinflector.kt is adapted from this project (which in turn credits Jiten). Also a great
reference for Android Yomitan-dictionary parsing.
- Jiten — deinflection rule data used by the deinflector.
- meikiocr by rtr46 (code Apache-2.0, models
meiki.text.detect.v0 and
meiki.txt.recognition.v0, LGPL-3.0) —
the default OCR engine; the models are downloaded from Hugging Face at runtime and the
pre/post-processing in
MeikiDecoder.kt is ported from its reference pipeline. The APK ships
modified graphs of both models (shorter detector inputs, any-width recognizer; LGPL-3.0, built
by tools/meiki-variants from the originals, whose weights they read at runtime).
- meikidroid by rtr46 — Android game OCR overlay that
served as a reference for the region-capture + overlay workflow and the hotkey idea.
- ONNX Runtime (MIT) — runs meikiocr and PaddleOCR on device.
- PaddleOCR (Apache-2.0) — PP-OCRv5 detector and
PP-OCRv6 recognizer ONNX models from ,
repackaged unmodified in the optional English OCR pack (GitHub Releases ).
Droid Sentence Miner is not affiliated with Anki, AnkiDroid, Yomitan, Jiten or GameSentenceMiner.
License
GPL-3.0. The adapted Yomihon deinflector remains available under its Apache-2.0 license
wherever required; see the file header for details.
