A Windows TSF (Text Services Framework) input method that lets you type Japanese on a Korean 2-beolsik keyboard layout. You type Korean syllables (와타시노, 갓코우, 한바아가아…) and KorJpnIme converts them to Japanese kana, then offers kanji candidates with a Mozc-quality viterbi-based segmentation engine.

🌏 Languages: English · 한국어 · 日本語

와타시  +  Space   →   私 / 渡し / 渡 / ワタシ / わたし
와타시노 + Space   →   私の / 渡しの / わたしの / ワタシノ
갓코우  +  Space   →   学校 / 月光 / がっこう / ガッコウ
ㅇㅗㅗ  (no Space) →   を       (object particle)
ㅇㅘㅏ  (no Space) →   は       (topic particle)
ㅇㅔㅔ  (no Space) →   へ       (direction particle)
한바아가아 + Space →   ハンバーガー (auto-katakana fallback)

No mode toggling, no romaji shim, no extra hotkeys to remember. The candidate window pops up next to your caret like every other JP IME on the platform.

• Korean 2-beolsik layout → Japanese kana with full compound vowel (ㅗ+ㅏ=ㅘ → わ) and compound jong (ㄴ+ㅈ=ㄵ → with proper migration) handling.

• Hatsuon ん / sokuon っ auto-detected from Korean batchim patterns (ㄴㅁㅇ → ん, ㅎ → っ, ㅅㅆ at terminal or before consonants → っ, etc.).

• Mozc OSS dictionary (~750k unique kana keys, ~1.29 M entries) for kanji candidates with proper Mozc cost values.

• Viterbi top-K segmentation powered by Mozc's bigram connection cost matrix. Multi-word phrases like わたしの decompose into segmented candidates (私の, 渡しの, ...) at the top of the candidate list.

• Verb conjugation support via Mozc's suffix table — たべます, よみました, きれいです segment correctly even when the full inflected form is not in the main dictionary.

• Auto-katakana fallback for loanwords whose dictionary path doesn't reach kanji (ハンバーガー for 한바아가아).

• Three Japanese-particle keystroke patterns that have no native Korean meaning, repurposed:

  • ㅇ-ㅗ-ㅗ → を (object, "wo" sound)
  • ㅇ-ㅘ-ㅏ → は (topic, "wa" sound)
  • ㅇ-ㅔ-ㅔ → へ (direction, "e" sound)

  Real long-vowel forms (おお, ええ) are unaffected because they require an explicit ㅇ between the vowels.

• Per-user learning dictionary — every kanji you select bubbles to the top of the list next time. Capped at 5000 entries by default (LFU eviction); knob in settings.ini.

• Settings hot-reload — edits to %APPDATA%\KorJpnIme\settings.ini apply on the next keystroke. No log-out required.

• Persistent katakana toggle (configurable, default RAlt+K) forces katakana output for the next sequence of keys.

• Custom display attribute — preedit text rendered with a dotted blue underline matching MS-IME / Mozc visual style.

1. Download the latest KorJpnIme-vX.Y.Z.zip from the GitHub release page (or build from source — see below).

2. Unzip anywhere.

3. From an admin PowerShell, run:

   powershell -ExecutionPolicy Bypass -File install.ps1

   The script self-elevates if needed. It copies files to C:\Program Files\KorJpnIme\, registers the COM server, and imports the TSF profile registry entries.

4. Log out and log back in. ctfmon caches IME profiles per session and only picks up new ones cleanly across a fresh session.

5. Settings → Time & Language → Language → Korean → Language options → Add a keyboard → Korean-Japanese IME.

6. Press Win+Space to switch to it whenever you want to type Japanese.

powershell -ExecutionPolicy Bypass -File "C:\Program Files\KorJpnIme\uninstall.ps1"
# add  -RemoveFiles  to also delete the install directory

Settings live at %APPDATA%\KorJpnIme\settings.ini (auto-created on first run with the documented defaults). Edit, save, and the IME picks up changes on the next keystroke.

[Hotkeys]
KatakanaToggle = RAlt+K          ; modifiers: Ctrl Shift Alt LAlt RAlt Win

[Behavior]
FullWidthAscii      = true       ; type 1 → １ etc. while IME is on
UserDictLearn       = true       ; remember every kanji you pick
UserDictMaxEntries  = 5000       ; LFU eviction cap; 0 = unlimited

• Windows 10/11
• Visual Studio 2022 Build Tools (MSVC v143 + Windows 10 SDK)
• Python 3.10+ (only for the data builder, not for runtime)

:: 1. Pull Mozc OSS data sources (~95 MB of TSV)
cd dict\mozc_src
fetch.sh           :: bash; or run inside WSL

:: 2. Build the binary engine inputs (~71 MB output)
cd ..
python build_viterbi_data.py

:: 3. Build the DLL + bundle a release folder
cd ..
tools\make_release.bat

Output: C:\Temp\KorJpnIme_release\ with the DLL, dictionaries, .reg files, install/uninstall scripts, and LICENSES.txt. Zip it up and ship.

tools\build_tests.bat

Builds and runs the 56-case unit suite (HangulComposer, BatchimLookup, Viterbi smoke). Exit code 0 iff every test passed.

mapping/syllables.yaml        Korean syllable → kana table (~150 entries)
   ↓ tsf/tools/gen_table.py
tsf/generated/
   ├── mapping_table.h        constexpr sorted Entry[] for binary search
   └── batchim_rules.h        받침 → kana suffix rules (sokuon / hatsuon)

dict/
   ├── jpn_dict.txt           legacy text dict (kana → kanji TSV, 18 MB)
   ├── kj_dict.bin            rich dict with lid/rid/cost (binary, 57 MB)
   ├── kj_conn.bin            Mozc bigram cost matrix (binary, 14 MB)
   ├── LICENSES.txt           NAIST IPAdic + Mozc + Okinawa attributions
   ├── build_viterbi_data.py  builds kj_*.bin from Mozc OSS sources
   └── build_dict_mozc.py     builds jpn_dict.txt from same sources

tsf/src/
   ├── dllmain.cpp            DLL entry, IClassFactory, registration
   ├── KorJpnIme.cpp          ITfTextInputProcessor + ITfDisplayAttributeProvider
   ├── KeyHandler.cpp         ITfKeyEventSink + viterbi-driven candidate build
   ├── HangulComposer.cpp     pure 2-beolsik state machine (testable in isolation)
   ├── Composition.cpp        preedit + commit via TSF edit sessions
   ├── DisplayAttributes.cpp  ITfDisplayAttributeInfo (dotted blue underline)
   ├── BatchimLookup.h        single-syllable Korean → kana with batchim rules
   ├── Dictionary.cpp         legacy text dict reader (mmap'd jpn_dict.txt)
   ├── RichDictionary.cpp     binary dict reader (mmap'd kj_dict.bin)
   ├── Connector.cpp          binary connection cost matrix (mmap'd kj_conn.bin)
   ├── Viterbi.cpp            top-K viterbi over RichDictionary + Connector
   ├── UserDict.cpp           per-user learning dict + LFU pruning
   ├── Settings.cpp           %APPDATA% settings.ini + hot-reload
   ├── CandidateWindow.cpp    Win32 popup, mouse + keyboard navigation
   └── KanaConv.h             hiragana ↔ katakana conversion helpers

tsf/tests/                    minimal header-only test runner (56 cases)
tools/                        install / uninstall / make_release / build_tests

• Reconversion not implemented. Selecting an already-committed string and asking for a different conversion (the standard ITfFnReconversion flow) is on the roadmap. Workaround: delete and retype.
• First install + every update needs a logout/login cycle. ctfmon caches the loaded TIP DLL per session; new DLLs are only picked up across a fresh login. Working as designed for TSF.
• Tested on Windows 11. Should work on Windows 10 (same TSF surface) but is not exercised there in CI.
• UserDict learning influences first-segment ranking only. Viterbi paths themselves don't yet incorporate user-chosen surfaces into their cost. Planned for a future patch.
• Top-K viterbi is K=5. Hard-coded; not yet a settings knob.
• No tray icon. Mode state is visible only via the preedit / candidate window.

This project is MIT-licensed (see LICENSE).

The shipped dictionary data (jpn_dict.txt, kj_dict.bin, kj_conn.bin) is derived from Google Mozc OSS (BSD-3) which itself derives from NAIST IPAdic and the Okinawa Public-Domain Dictionary. Full third-party attribution text is in dict/LICENSES.txt and MUST be redistributed alongside the data.

Issues and pull requests welcome. For significant changes, please open an issue first to discuss what you'd like to change.

When fixing bugs, the existing test suite (tools\build_tests.bat) should keep passing; add a regression case for the bug you're fixing whenever it's testable in pure-logic land.