Live conversion
Pause briefly after typing hiragana and the top candidate appears automatically. See a possible result before pressing Space and continue refining it as the sentence develops.
Experimental Japanese IME / Windows / Rust
rakukan
Give Japanese input
the context it needs.
rakukan brings the LLM-based conversion engine at the heart of karukan to Japanese input on Windows. It combines a small local LLM with Mozc-family dictionaries to explore a different approach to conversion. Live conversion anticipates candidates while you type, and local learning adjusts their order over time. My sincere thanks go to Hitoshi Togasaki for creating and sharing karukan.
きょうはいいてんきです
今日はいい天気です
- 1今日はいい天気です
- 2今日は良い天気です
- 3きょうはいい天気です
01 / A different conversion
Dictionary reliability,
with contextual language modeling.
Traditional kana-to-kanji conversion builds candidates from readings, dictionaries, frequency, and learned choices. rakukan adds a compact language model that can consider the sentence around the reading while keeping literal numbers and Latin characters intact.
The text is not sent to a cloud AI. The model, dictionary, and conversion history run on the local machine, making rakukan an experiment in context-aware conversion that remains under the user's control.
02 / Input features
Let candidates develop
as the sentence grows.
Range- and punctuation-aware conversion
Adjust the conversion range with Shift and the arrow keys. Sentences containing punctuation or brackets are divided into blocks that can be converted and confirmed in sequence.
Literal protection
Years, model numbers, Latin text, out-of-vocabulary symbols, and emoji are protected from disappearing during conversion. F6 through F10 provide familiar script and width conversions.
Dictionary and learning
Mozc-family dictionary candidates complement the LLM. Confirmed results feed the user dictionary and learning history, allowing personal vocabulary and choices to influence later rankings.
03 / Out-of-process design
Keep the IME light.
Move inference elsewhere.
A lightweight TSF client lives inside the Windows input framework, while rakukan-engine-host.exe manages the LLM and GPU backends. Keeping heavy inference outside the IME process reduces how much the input component must carry.
CPU, Vulkan, and CUDA engine DLLs can be selected in configuration. The model variant and number of GPU-offloaded layers can be tuned for the machine and applied by switching IME mode.
Windows appstext fields
→TSF clientlightweight DLL
⇄Engine hostLLM + dictionary
Prototype the next input
Test a different future
for Japanese IMEs.
Review the warnings, installation steps, and known issues in the README before trying it.
Explore rakukan on GitHub ↗