romanizer

Fast, CPU-only romanization of names and text from any script into plain Latin letters.

from romanizer import romanize

romanize("Володимир Зеленський")  # 'Volodymyr Zelenskyi'
romanize("Александр Пушкин")  # 'Aleksandr Pushkin'
romanize("Γιώργος Παπανδρέου")  # 'Giorgos Papandreou'
romanize("محمد صلاح")  # 'Mohamed Salah'
romanize("习近平")  # 'Xi Jinping'
romanize("김정은")  # 'Kim Jeongeun'
romanize("नरेन्द्र मोदी")  # 'Narendra Modi'
romanize("Nguyễn Văn Đức")  # 'Nguyen Van Duc'
romanize("Zeleнський in Київ")  # 'Zelenskyi in Kyiv'

The output is always printable ASCII. No apostrophes standing in for soft signs, no backticks for ayin, no stray spaces before commas, no emoji, no №, no €.

Why another one

Single-library solutions all fail somewhere:

• unidecode / anyascii work character by character. Ukrainian comes out Russian-flavoured (Volodimir, Kiyiv), soft signs become ', Kazakh gets @, Japanese kanji are read as Chinese (東京 → Dong Jing), and CJK output has stray spaces.
• uroman is good but 50–100× slower than everything else (~0.5 ms per name, 1 s to load).
• Dedicated libraries (pypinyin, pykakasi, korean-romanizer, arabic-name-romanizer) are excellent for their one script and know nothing about the others or about mixed input.

romanizer splits input into same-script runs, guesses the language inside a script (Ukrainian vs Russian vs Bulgarian, Japanese vs Chinese, Persian vs Arabic), routes each run to the best engine, and passes everything through one final ASCII-only layer.

Installation

pip install romanizer                 # Cyrillic, Greek, Latin, Armenian, Georgian, Ethiopic
pip install "romanizer[arabic]"       # + Arabic, Persian, Dari, Pashto, Urdu (arabic-name-romanizer, ONNX CPU)
pip install "romanizer[cjk]"          # + Chinese, Japanese, Korean
pip install "romanizer[icu]"          # + Indic and South-East Asian scripts (needs libicu)
pip install "romanizer[thai]"         # + Thai (pythainlp, rule-based engine only)
pip install "romanizer[all]"

Nothing here touches a GPU. Missing extras never crash: the affected script falls back to a readable approximation and is flagged as unsupported in analyze().

Language support

Status: ✅ supported · ⚠️ supported with caveats · ❌ not supported for now.

Measured agreement with Wikidata

scripts/evaluate_wikidata.py romanizes the native-language labels of about 1,400 people per language from Wikidata and compares them with the English labels. Token recall is the share of English-label words that appear verbatim in our output. English labels are not gold transliterations: they follow anglicised habits (Vladimir for Володимир), nicknames and drop patronymics, so these are floors for comparing languages, not accuracies.

Known ambiguities

• Han without kana is read as Chinese. 山田太郎 alone gives Shan Tian Tai Lang; pass lang="ja" when you know the source is Japanese.
• Cyrillic without language-specific letters is read as Russian. Володимир has no Ukrainian-only letters and becomes Volodimir; with a surname (Зеленський) or lang="uk" it becomes Volodymyr. Likewise Bulgarian is only recognised by a vocalic ъ (Търново), so Христо Ботев needs lang="bg" to become Hristo Botev instead of Khristo Botev. When you know the language, pass it.
• Names mode splits Chinese and Korean tokens that start with a common surname. 田中 (Tanaka) is a Chinese surname too, so without lang="ja" it becomes Tian Zhong; 王府 becomes Wang Fu. Use mode="text" for running text.
• Serbian written in Latin letters loses Đ → D (Đorđe → Dorde) while the Cyrillic form gives Djordje: Đ is also Vietnamese, where D is right, and nothing in the text says which.
• Dari and Iranian Persian share an alphabet. Both are detected as fa; pass arabic_country="AF" for Afghan data to prefer Ahmad Shah Massoud over Ahmed Shah Masoud.
• Hindi schwa deletion is a heuristic. It gets Tendulkar, Kolkata, Arvind right and Draupadi wrong (Draupdi). Use overrides for the names that matter.

API

from romanizer import Romanizer, romanize, romanize_many, analyze

romanize("Київ")  # auto-detect
romanize("Володимир", lang="uk")  # language hint
romanize_many(["محمد", "أحمد", "Київ"])  # batches Arabic runs into one model call (20× faster)

r = Romanizer(
    mode="names",  # or "text": no surname splitting, no title-casing
    unsupported="fallback",  # or "raise" / "drop" for scripts with no engine (Hebrew)
    arabic_country="EG",  # regional spelling preference for Arabic names
    overrides={  # exact spellings that win over every engine, whole input or any token
        "Зеленський": "Zelenskyy",
        "محمد صلاح": "Mo Salah",
    },
)
r.romanize("Володимир Зеленський")  # 'Volodymyr Zelenskyy'

a = analyze("Zeleнський in Київ")
a.output  # 'Zelenskyi in Kyiv'
a.segments[1].script  # 'Cyrillic'
a.segments[1].lang  # 'uk'
a.segments[1].engine  # 'table-cyrillic'
a.segments[1].supported  # True

Command line:

romanize "Володимир Зеленський" "محمد صلاح"
romanize --lang uk < names.txt > names_latin.txt
romanize --overrides known_spellings.tsv < names.txt      # 'source<TAB>latin' per line
romanize --analyze "Zeleнський in Київ"

Speed

Single thread, Apple M-series, warm engines, microseconds per string:

Ten million mixed records take minutes on one core. Always use romanize_many for Arabic.

Output guarantees

1. Only printable ASCII leaves the library. Every character is either mapped by a table, decomposed by NFKD with its marks removed, mapped by an explicit table for letters NFKD cannot decompose, or, as a last resort, transliterated by anyascii.
2. Backticks, carets and tildes never survive: they only ever come from transliteration schemes.
3. Whitespace is collapsed to single spaces and trimmed.
4. Casing of cased scripts (Cyrillic, Greek, Latin) is preserved: ЩЕРБАК → SHCHERBAK, Щербак → Shcherbak. Caseless scripts are title-cased in names mode.

Contributing

uv sync
uv run pytest          # 100% line and branch coverage is enforced
uv run ruff check .
uv run mypy            # strict

See CONTRIBUTING.md for the layout, the checklist for adding a language, and the pull request expectations. Security reports: SECURITY.md. Changes: CHANGELOG.md.

License

MIT for this package (see LICENSE; dependency licenses in THIRD_PARTY_NOTICES.md). Optional engines keep their own licenses: anyascii (ISC), pypinyin (MIT), PyICU (MIT), pythainlp (Apache-2.0), arabic-name-romanizer (MIT), and in the [cjk] extra pykakasi and korean-romanizer, which are GPL-3.0-or-later. If GPL is a problem for your deployment, skip the [cjk] extra: Chinese, Japanese and Korean then fall back to anyascii.