num2words2 library - Convert numbers to words in multiple languages

num2words2 converts numbers like 42 to words like forty-two across 120+ languages and 170+ locale codes (see REFERENCE.md for the full table), producing cardinals, ordinals, currency, year, fractions, bank-cheque format, and aviation/ICAO digit-by-digit phraseology.

It is a Rust port of the num2words conversion engine with a thin Python binder. Every conversion runs in a compiled Rust core (PyO3/abi3); Python only normalises the arguments and shapes the result. The output is byte-for-byte identical to the original pure-Python library — validated against a frozen ~150,000-case corpus — while running typically 4–12× faster, and up to 26× on some languages and 150× on string parsing (see Performance).

The project is hosted on GitHub, and the full documentation is available in the Wiki. Contributions are welcome.

Performance

Because the conversion engine is compiled Rust, num2words2 is several times faster than the original pure-Python num2words. Benchmarked against num2words 0.5.14, Apple M-series, nanoseconds per call:

Operation	original	num2words2	speedup
int → cardinal (8765)	24,384 ns	2,099 ns	11.6×
float → cardinal (1234.56)	33,107 ns	3,367 ns	9.8×
ordinal	24,969 ns	2,407 ns	10.4×
currency	32,518 ns	3,345 ns	9.7×
year	13,484 ns	1,434 ns	9.4×
string input ("1234")	335,862 ns	2,200 ns	152.6×
French cardinal	61,046 ns	2,359 ns	25.9×
German cardinal	63,046 ns	2,386 ns	26.4×
Russian cardinal	4,781 ns	1,001 ns	4.8×

The speedup is largest where the original does heavy Python-level parsing (string inputs) or deep recursion (French, German); it never comes at the cost of accuracy — output is byte-for-byte identical to the original across a frozen ~150,000-case corpus of cardinals, ordinals, currency, years, fractions, floats and strings.

Installation

Install from PyPI (Python 3.10+) — the wheel bundles the Python binder and the compiled Rust extension, built per platform:

pip install num2words2

Prebuilt wheels are published for:

• Linux — x86_64 and aarch64 (manylinux2014)
• macOS — Apple Silicon (arm64); Intel (x86_64) from a later release
• Windows — x86_64

On any other platform, pip falls back to the source distribution and builds the extension locally — this needs a stable Rust toolchain (rustup).

Slim vs. full wheels

num2words2 ships in two flavours:

• slim (the PyPI default, ~2 MB) — everything except native language-detection. num2words_sentence(lang=None) auto-detect is disabled; pass an explicit lang and every converter works.

• full (~140 MB) — embeds the lingua-rs models so num2words_sentence(lang=None) auto-detects the language natively. Full wheels are attached to each GitHub Releases entry; install one directly:

  pip install https://github.com/jqueguiner/num2words2/releases/download/v1.0.18/num2words2-1.0.18-cp38-abi3-<platform>.whl
  

The compiled extension is distributed only through these wheels (it is not committed to the repository).

To build from source you need a stable Rust toolchain and maturin:

maturin build --release                        # full (bundles lingua models)
maturin build --release --no-default-features  # slim (~2 MB, no models)

Development Setup

The project uses pre-commit hooks to ensure code quality. To set up your development environment:

# Install pre-commit
pip install pre-commit

# Install the git hook scripts
pre-commit install

# Run hooks on all files (optional, useful for initial setup)
pre-commit run --all-files

This will automatically format and lint your code before each commit using:

• autopep8 - PEP 8 formatting
• autoflake - removes unused imports and variables
• isort - sorts imports
• flake8 - style and quality checks
• trailing-whitespace removal
• end-of-file fixing

Testing

The library uses pytest for testing. First, install the development dependencies:

make dev-install

Then, you can run the test suite using several methods:

• Run basic tests: This runs tests with your current Python environment.

  make test

• Run with Tox: This runs tests against all supported Python versions, which is the standard for CI.

  tox

Generating End-to-End Tests with LLMs

The repository includes a powerful script to generate high-quality, realistic test cases using Large Language Models (LLMs). This helps ensure accuracy across multiple languages and complex scenarios.

What it does: The tests/scripts/generate_llm_tests.py script uses an LLM (like GPT-4o) to create sentences containing numbers, dates, and currencies, and then generates the expected word-for-word conversion.

Requirements:

• An OpenAI API key. You must set it as an environment variable: export OPENAI_API_KEY='your-key-here'

How to Use:

To generate 10 new test sentences for French and Spanish, you can run:

python tests/scripts/generate_llm_tests.py --languages fr,es --samples 10

The new tests will be appended to tests/data/e2e_test_sentences.csv.

Key Options:

• --languages: Comma-separated list of language codes (e.g., en_IN,de,it).
• --samples: Number of samples to generate per language.
• --mode: Use sentences for full sentences or numbers for direct number-to-word conversions.
• --model: The OpenAI model to use (e.g., gpt-4o, gpt-4o-mini).
• --output: Specify a different output file.
• --overwrite: Overwrite the output file instead of appending.

This tool is essential for expanding test coverage and ensuring the library's robustness.

Usage

Command line:

$ num2words2 10001
ten thousand and one
$ num2words2 24,120.10
twenty-four thousand, one hundred and twenty point one
$ num2words2 24,120.10 -l es
veinticuatro mil ciento veinte punto uno
$ num2words2 2.14 -l es --to currency
dos euros con catorce céntimos

In code there's only one function to use:

>>> from num2words2 import num2words
>>> num2words(42)
forty-two
>>> num2words(42, to='ordinal')
forty-second
>>> num2words(42, lang='fr')
quarante-deux

Besides the numerical argument, there are two main optional arguments, to: and lang:

to: The converter to use. Supported values are:

• cardinal (default)
• ordinal
• ordinal_num
• year
• currency

lang: The language in which to convert the number. Supported values are:

• en (English, default)
• am (Amharic)
• ar (Arabic)
• az (Azerbaijani)
• be (Belarusian)
• bn (Bangladeshi)
• ca (Catalan)
• ce (Chechen)
• cs (Czech)
• cy (Welsh)
• da (Danish)
• de (German)
• en_GB (English - Great Britain)
• en_IN (English - India)
• en_NG (English - Nigeria)
• es (Spanish)
• es_CO (Spanish - Colombia)
• es_CR (Spanish - Costa Rica)
• es_GT (Spanish - Guatemala)
• es_VE (Spanish - Venezuela)
• eu (EURO)
• fa (Farsi)
• fi (Finnish)
• fr (French)
• fr_BE (French - Belgium)
• fr_CH (French - Switzerland)
• fr_DZ (French - Algeria)
• he (Hebrew)
• hi (Hindi)
• hu (Hungarian)
• hy (Armenian)
• id (Indonesian)
• is (Icelandic)
• it (Italian)
• ja (Japanese)
• kn (Kannada)
• ko (Korean)
• kz (Kazakh)
• mn (Mongolian)
• lt (Lithuanian)
• lv (Latvian)
• nl (Dutch)
• no (Norwegian)
• pl (Polish)
• pt (Portuguese)
• pt_BR (Portuguese - Brazilian)
• ro (Romanian)
• ru (Russian)
• sl (Slovene)
• sk (Slovak)
• sr (Serbian)
• sv (Swedish)
• te (Telugu)
• tet (Tetum)
• tg (Tajik)
• tr (Turkish)
• th (Thai)
• uk (Ukrainian)
• vi (Vietnamese)
• zh (Chinese - Traditional)
• zh_CN (Chinese - Simplified / Mainland China)
• zh_TW (Chinese - Traditional / Taiwan)
• zh_HK (Chinese - Traditional / Hong Kong)

You can supply values like fr_FR; if the country doesn't exist but the language does, the code will fall back to the base language (i.e. fr). If you supply an unsupported language, NotImplementedError is raised. Therefore, if you want to call num2words with a fallback, you can do:

try:
    return num2words(42, lang=mylang)
except NotImplementedError:
    return num2words(42, lang='en')

Additionally, some converters and languages support other optional arguments that are needed to make the converter useful in practice.

Wiki

For the full documentation, including installation, API details, CLI usage, supported locales, sentence conversion, currency handling, and migration guidance, please check the Wiki. Feel free to propose wiki enhancements.

History

num2words is based on an old library, pynum2word, created by Taro Ogawa in 2003. Unfortunately, the library stopped being maintained and the author can't be reached. There was another developer, Marius Grigaitis, who in 2011 added Lithuanian support, but didn't take over maintenance of the project.

Virgil Dupras from Savoir-faire Linux based himself on Marius Grigaitis' improvements and re-published pynum2word as num2words.

Relationship to num2words

num2words2 is not a divergent fork. It is a Rust port of the num2words conversion engine with a Python binder, kept output-compatible with the original.

The original project's issues and pull requests are actively monitored. Valid, relevant fixes and features are reviewed and then ported into this project's Rust core, with byte-parity re-verified against the frozen corpus. num2words2 therefore tracks upstream improvements while delivering them natively — and several times faster.

Jean-Louis Queguiner