Language Packs

Add language dictionaries by importing language pack files. Keyboard layouts are configured separately.

Quick Summary

What	Description
Purpose	Add language support (dictionary + predictions)
Access	Settings > 🌐 Multi-Language > Import Pack
Contents	Dictionary, frequency list, contractions

[!TIP] Greek, Russian, German, French, Spanish, Italian, Portuguese, Dutch, Swedish, Turkish, Indonesian, Malay, Tagalog, Swahili and more are already built — you don’t need to build anything. See Prebuilt Language Packs below, then import via Settings > 🌐 Multi-Language > Import Pack.

What’s in a Language Pack

Each language pack includes:

Component	Description
Dictionary	Binary word list for predictions
Contractions	Language-specific contractions (if available)
Frequency list	Word frequencies for the dictionary
Notice	Optional licensing and attribution text
Swipe model	Optional declared model; runtime approval is checked separately

Import Limits

Packs can contain up to 100,000 dictionary words. The dictionary is limited to 16 MiB, an optional model to 8 MiB, and the archive to 64 entries and 64 MiB of extracted data. CleverKeys checks the dictionary header and extracted bytes, so a smaller advertised count does not bypass these limits. An oversized update is rejected before replacing the existing dictionary.

A dictionary import does not automatically make every script swipe-compatible. Bangla National and Provat tap layouts are available; Bangla transliteration and a validated Bangla swipe dictionary/model remain outstanding.

Bundled Languages

CleverKeys bundles these dictionaries in the app (no import needed): English, Spanish, French, German, Italian, Portuguese, Swedish.

Every other language is added by importing a language pack (below).

Prebuilt Language Packs

These packs are already built and ship in the repository under the langpacks release — download the .zip and import it (no need to build anything):

Language	Pack file	Words
Greek (Ελληνικά)	`langpack-el.zip`	39,860
Russian (Русский)	`langpack-ru.zip`	50,000
German (Deutsch)	`langpack-de.zip`	40,000
Spanish (Español)	`langpack-es.zip`	50,000
French (Français)	`langpack-fr.zip`	40,000
Italian (Italiano)	`langpack-it.zip`	40,000
Portuguese (Português)	`langpack-pt.zip`	40,000
Dutch (Nederlands)	`langpack-nl.zip`	40,000
Swedish (Svenska)	`langpack-sv.zip`	40,000
Turkish (Türkçe)	`langpack-tr.zip`	40,000
Indonesian	`langpack-id.zip`	28,637
Malay	`langpack-ms.zip`	25,861
Tagalog/Filipino	`langpack-tl.zip`	27,922
Swahili	`langpack-sw.zip`	20,000
Ukrainian (Українська)	`langpack-uk.zip`	50,000
Bulgarian (Български)	`langpack-bg.zip`	35,027
Macedonian (Македонски)	`langpack-mk.zip`	50,000
Hebrew (עברית)	`langpack-he.zip`	50,000

Most packs are generated from the wordfreq corpus (real word frequencies), most of them validated against the AOSP LatinIME word list; Swahili comes from the Swwiki Swahili-Wikipedia list. See Attribution. The bundled English dictionary is different: it’s built by a dedicated evidence-classification pipeline (98,140 words validated against multiple spelling oracles, with typo and junk filtering) rather than a raw frequency cut.

Importing Language Packs

Step 1: Obtain the Language Pack

Prebuilt (recommended): download a .zip from the langpacks release (e.g. langpack-el.zip for Greek) to your device.
Build your own for a language not listed (see below).

Step 2: Import via Multi-Language

Open Settings
Go to the 🌐 Multi-Language section
Tap Import Pack
Choose the language pack .zip file
The pack is extracted and installed (you’ll see “Installed: N language pack(s)”)

Step 3: Select the Language

After import the language becomes selectable immediately:

In Settings > 🌐 Multi-Language, set Primary Language (or Secondary Language) to the imported language — e.g. Greek (Ελληνικά).
The dictionary loads automatically for predictions and autocorrect.
Pick the matching keyboard layout (e.g. the Greek layout) if you want the native script on the keys.

Building Custom Language Packs

For languages not bundled, you can create your own using the provided Python scripts:

Requirements

Python 3.x
wordfreq package (optional, for frequency data)

Using Build Scripts

# Navigate to scripts directory
cd scripts/

# Install prerequisite
pip install wordfreq

# Option 1: Two-step build from wordfreq (any language wordfreq supports)
python get_wordlist.py --lang sv --output sv_words.txt --count 50000
python build_langpack.py --lang sv --name "Swedish" --input sv_words.txt --use-wordfreq --output langpack-sv.zip

# Option 2: Build from pre-existing binary dictionary (.bin file)
python build_langpack.py --lang sv --name "Swedish" --dict ../src/main/assets/dictionaries/sv_enhanced.bin --output langpack-sv.zip

# Option 3: Build from your own word list: one entry per line, either "word",
# "word,frequency", "word<TAB>frequency" or "word frequency" (integer or decimal).
# The --input of build_langpack.py is the same kind of word list; raw prose is not tokenized.
python build_dictionary.py --input my_words.csv --output custom.bin
python build_langpack.py --lang xx --name "MyLang" --dict custom.bin --output langpack-xx.zip

Scripts Available

Script	Purpose
`build_langpack.py`	Create .zip language pack (with licence metadata + `NOTICE.txt`)
`build_dictionary.py`	Build binary dictionary from a word list (word / CSV / tab / space formats)
`build_all_languages.py`	Batch build all supported languages
`get_wordlist.py`	Extract top N words from wordfreq

Language Pack Structure

langpack-{lang}.zip
├── manifest.json          # Metadata: code, name, version, wordCount, model
├── dictionary.bin         # V2 binary dictionary (required)
├── unigrams.txt           # Word-frequency list for language detection
├── contractions.json      # Language contractions (optional)
├── prefix_boost.bin       # Aho-Corasick prefix trie (optional, non-English)
├── model.onnx             # CTC swipe encoder (optional; the six non-Latin scripts)
└── NOTICE.txt             # Attribution, licence and change note for the word data

Since 2026-09-26 every prebuilt pack also records its licence in manifest.json — "license" (GPL-3.0-only), "attribution" (credit line + upstream licence + change note) and "source" (upstream URL) — so the credits travel with a pack downloaded on its own. The importer keeps both on the device: the installed manifest holds the three keys and NOTICE.txt is installed beside the dictionary, and the language-pack manager shows them (see Attribution). All three keys and NOTICE.txt are optional — a pack without them still imports.

manifest.json and dictionary.bin are required; the importer rejects a pack missing either, or a dictionary.bin without the V2 (CKDT) header.

The `model.onnx` member (ru, el, uk, bg, mk, he)

The six non-Latin-script packs carry their own CTC swipe encoder. Those languages cannot swipe-decode without their pack in the first place, so from 2026-09-10 the model travels with the pack instead of sitting in the APK — ~3 MB every other user was carrying for nothing.

A pack that ships a model declares it:

"model": { "file": "model.onnx", "sha256": "8fffa75c…" }

Two separate checks apply, and they answer different questions:

On import the bytes must hash to what the manifest says. This catches a corrupt or truncated download and fails the import with that reason, rather than installing a language whose swipe silently never works. A model.onnx with no manifest entry is skipped (nothing can verify it), and anything over 8 MiB is rejected outright.
On load the app compares the model against a SHA-256 compiled into the app and loads it only on byte-identity. It trusts nothing the manifest says — a pack’s manifest is written by whoever built the pack. A mismatch is treated as “no model”: swipe falls back to the geometric engine and the reason is logged.

So a hand-built pack can carry any model.onnx it likes and the app will still only ever run the encoder it shipped a hash for. The practical consequence: a new script model needs an app update as well as a pack, by design.

Languages Supported by wordfreq

Languages available through the wordfreq Python package:

European: Swedish (sv), Norwegian (nb), Danish (da), Finnish (fi), Polish (pl), Czech (cs), German (de), French (fr), Spanish (es), Italian (it), Portuguese (pt)
Asian: Japanese (ja), Korean (ko), Chinese (zh)
Other: Russian (ru), Arabic (ar), Hebrew (he), Hindi (hi), Turkish (tr)

Managing Language Packs

View Installed

Go to Settings > Multi-Language (with multi-language enabled)
Under Language Packs, tap Manage
Each installed pack is listed with its code and word count; tap Source & license on a pack to see where its word data came from (see Attribution)

Remove a Language Pack

Open Manage as above
Tap Delete on the pack

Offline Operation

Once imported, all language features work offline:

Feature	Works Offline
Typing	✅
Predictions	✅
Autocorrect	✅
Contractions	✅

Tips and Tricks

Start with English: English is fully bundled and ready to use
Build for your language: Use the scripts to create packs for unsupported languages
Share packs: Language pack files can be shared with other users
Backup first: Export your settings before major imports

Common Questions

Q: How do I add a new language?

A: If it’s in the Prebuilt Language Packs list (Greek, German, French, Spanish, Italian, Portuguese, Dutch, Swedish, Turkish, and more), just download the .zip and import it via Settings > 🌐 Multi-Language > Import Pack — no building required. For an unlisted language, build a pack with the Python scripts, then import the same way.

Q: How do I add Greek?

A: Download langpack-el.zip from the langpacks release, then Settings > 🌐 Multi-Language > Import Pack, and set Primary/Secondary Language to Greek (Ελληνικά). Greek word suggestions then work offline.

Q: How do I add Russian?

A: Same flow — download langpack-ru.zip from the langpacks release, Import Pack, set language to Russian (Русский), and pick the Cyrillic (ЙЦУКЕН) layout. ~50,000 words, works offline.

Q: Can I use a language without importing a pack?

A: Basic typing works with any layout, but predictions and autocorrect require a dictionary.

Q: Why is my language not available?

A: Build it yourself using the build_langpack.py script with the wordfreq package.

Q: Do I need to download English?

A: No, English is included by default.

Attribution

Each pack’s word list is an adaptation (“converted to a CleverKeys frequency list”) of third-party data, and each zip carries a NOTICE.txt with the full credit:

Packs	Source	Licence
en, en-wordfreq, de, es, fr, it, nl, pt, sv, ru, el, tr, he, uk, bg, mk, id, ms, tl	wordfreq (Robyn Speer et al.) — incorporating SUBTLEX (Brysbaert et al.), OpenSubtitles, Wikipedia and Google Books Ngram data	CC BY-SA 4.0
de, es, fr, it, nl, pt, sv, ru, el, tr, he	AOSP LatinIME word lists, used as an inclusion oracle	Apache-2.0
en-opensubtitles, en-opensubtitles-50k	FrequencyWords 2018 (Hermit Dave), from OpenSubtitles.org via the OPUS corpus (Lison & Tiedemann, 2016)	CC BY-SA 4.0
sw	Swwiki (Kevin Donnelly), from Swahili Wikipedia	CC BY-SA 3.0, adapted under 4.0

The packs are distributed under GPL-3.0-only. CC BY-SA 4.0 adaptations may be licensed under GPLv3 (CC compatible licenses); Swwiki’s CC BY-SA 3.0 reaches 4.0 through its §4(b) later-version clause; Apache-2.0 is one-way compatible with GPLv3. The former en-norvig-50k pack was withdrawn on 2026-09-26 (Google Web 1T data, no redistribution grant). See the repository NOTICE file.

Seeing a pack’s attribution in the app

Settings > Multi-Language > Language Packs > Manage, then tap Source & license on a pack. It expands to show:

Pack license — the pack’s licence (GPL-3.0-only for every prebuilt pack)
the credit line from the manifest: upstream author, upstream licence and the change note
Source — the upstream URL(s); tap one to open it in the browser (only http/https addresses are tappable, since a pack’s contents come from whoever made it)
View full notice — the pack’s complete NOTICE.txt, as installed on the device

Packs built before 2026-09-27 carry none of this, and the section says so. Packs imported with app versions before the attribution viewer kept the manifest keys but not NOTICE.txt; View full notice then asks you to import the pack again. Re-importing the current zip from the langpacks release fixes both cases and does not change your settings.

Adding Layouts - Use downloaded layouts
Multi-Language - Type in multiple languages
Autocorrect - Per-language corrections

Technical Details

See Language Packs Technical Specification.

All docs Report doc issue