offline-translate

Offline machine translation in Python using the same on-device neural translation models that the Google Translate Android app downloads for offline use.

• ~100 languages, each pack translates to/from English (~50 MB per language)
• Fully offline once a pack is downloaded; no API key, no account
• Language pairs without English (e.g. Hebrew → French) pivot through English, like the app does
• Runs on CPU via LiteRT (TensorFlow Lite); ~0.3 s per sentence

pip install offline-translate

Supported platforms are those with ai-edge-litert wheels: macOS (Apple Silicon), Linux (x86_64, aarch64) and Windows (x64), Python 3.10+.

Usage

from offline_translate import Translator

t = Translator("~/.local/share/offline-translate")   # where models are stored

t.download("he")                          # Hebrew <-> English, ~55 MB
t.translate("Good morning!", target="he").text
# 'בוקר טוב!'

t.translate("בוקר טוב", target="en")      # source is auto-detected among installed languages
# Translation(text='Good morning', source='iw', target='en', via=('iw', 'en'))

Managing languages

for lang in t.languages():                # full catalog (fetched once from Google, then cached)
    print(lang.code, lang.name, lang.installed, lang.download_size, lang.update_available)

t.installed_languages()                   # ['en', 'iw'] — works offline, no catalog needed
t.download("fr", progress=lambda done, total: print(f"{done}/{total}"))
t.remove("fr")

t.update_catalog()                        # check Google for a newer catalog
[l for l in t.languages() if l.update_available]   # then t.download(code) to update

Language codes follow Google's catalog (iw for Hebrew, zh-Hant, jw, …). Common aliases (he, zh-TW, jv) and English names ("Hebrew") are accepted everywhere.

English is built in: there is no separate English pack, since every pack contains both directions to and from English.

API

Network errors surface as OSError.

Command line

The repository also has a small CLI built on the library (not part of the PyPI package):

./translate get he                 # download Hebrew
./translate "Good morning"         # English -> Hebrew by default; source auto-detected
./translate -s he -t fr "בוקר טוב"  # any pair, via English
./translate list --installed

tools/pull_from_phone.py copies already-downloaded packs from a rooted Android phone over adb.

How it works

The Android app reads a manifest at https://dl.google.com/translate/offline/manifest/gtr/prod/manifest.pb, which lists versions of a package profile — the catalog of language packs and their files. Each translation direction is a TFLite encoder, an LSTM decoder run step by step with beam search, and SentencePiece tokenizers.

Limitations

• 15 languages (am, az, bs, hy, kk, ky, ml, mn, ne, pa, ps, si, so, tg, uz) still ship an older model format with a WordPiece tokenizer that isn't supported yet; they're listed with supported=False.
• Quality matches the app's offline mode, which is below online Google Translate.

Development

pip install -e ".[test]"
pytest                          # offline unit tests (fake Google server, fake model)
pytest --network                # + checks against the real Google catalog
pytest --models=~/models        # + end-to-end translation with installed Hebrew models

Disclaimer

This project is not affiliated with or endorsed by Google. It contains no Google code or models; language packs are downloaded from Google's public servers at runtime and are subject to Google's terms.