Language notes

Per-language behaviour that goes beyond the support matrix in the README.

Germanic compound numbers (de, nl, da, sv, fy)

These languages write numbers as single compound words. The extractors split them into their component number words and evaluate the result, so both the compound and the spaced form parse:

Language	Example	Value
German	einundzwanzig, zweihundertdreiundvierzig	21, 243
Dutch	eenentwintig, tweehonderddrieëntwintig	21, 223
Danish	toogfyrre, nihundredenioghalvfems	42, 999
Swedish	tjugoen, etthundratjugotre, tvåtusen tjugotre	21, 123, 2023
West Frisian	ienentweintich, twahûnderttrijeentweintich	21, 223

Dutch diacritics (één, drieëntwintig) are normalized before parsing. West Frisian diacritics (sân, hûndert, tûzen) are likewise normalized, and the variant spellings without the etymological -s- (sechtjin, sechtich) are accepted alongside the standard sechstjin/sechstich. Danish accepts both the standard spellings (elleve, halvtreds, halvfjerds) and the legacy variants kept for backward compatibility.

Czech declensions

Hundreds and thousands decline: sto, dvě stě, tři sta, pět set, dva tisíce. Both pronouncing and extraction use the declined forms.

Basque vigesimal counting

Basque counts in twenties. The -ta joiner forms are understood when extracting: hogeita bat (20 + 1 = 21), berrogeita hamar (40 + 10 = 50), laurogeita hemeretzi (80 + 19 = 99).

Romance languages (pt, gl, mwl, ast, an)

Portuguese, Galician, Mirandese, Asturian and Aragonese share a declarative NumberVocabulary + RomanceNumberExtractor implementation in ovos_number_parser.util. Each language is described by its vocabulary (units, tens, hundreds, scales, ordinals, fractions, gender rules and joiner placement). The shared engine handles pronunciation and extraction for all of them. This is the preferred path for adding new Romance languages.

Portuguese distinguishes European and Brazilian variants (pt-PT reads 16 as dezasseis, pt-BR as dezesseis) and inflects um/uma, dois/duas and the hundreds (duzentos/duzentas) for grammatical gender.

Aragonese pronounces the general forms of the current academic norm (cuatre, ueito, deciséis, vintiún, trenta y cinco, cient) and accepts the dialectal variants (quatre, güeito, setse, vente, noranta) on extraction. Ordinals use the characteristic -eno series (cuatreno, cinqueno, onceno). Tens ordinals above the attested range follow the same productive -eno pattern.

Spanish and Catalan

es and ca handle fused twenties (veintiuno, vint-i-u), the irregular hundreds (quinientos, setecientos, novecientos), scale words including millón/millones and milió/milions, and additive continuations such as dos mil veintitrés.

English

extract_number understands the and connector inside a number (one hundred and one = 101), digit grouping (1,000,000) and spoken-style comma pauses (two thousand, twenty three).

Kabyle

Kabyle carries two numeral systems: the pan-Amazigh cardinals (yiwen, sin, kṛaḍ, ... mraw) and the Algerian-Arabic borrowed numerals used for everyday counting above ten (ḥḍac = 11, ɛecrin = 20, mya = 100, alef = 1000). Pronunciation uses the Amazigh forms for 0-10 and the Arabic-derived forms above ten, joining compounds with the conjunction u with the unit before the ten (waḥed u ɛecrin = 21). Extraction accepts both systems, feminine forms (yiwet, snat), additive Amazigh compounds (mraw d yiwen = 11) and diacritic-stripped spellings (hdac).

Ordinals use amezwaru (f. tamezwarut) for "first" and wis/tis + cardinal otherwise. The only attested fraction noun is azgen (half).

Coverage boundary: pronunciation spans 0–9999. The minus word, a decimal-separator word, ordinal suffix forms, larger scale words and further fraction nouns are not reliably attested and are deliberately omitted.

Persian

Numbers are written and parsed in Persian script (بیست و یک = 21). Both extraction and pronunciation cover units through millions, including decimal readings.

Known gaps

These are cases where the generic parity fallback (see below) is in use and its result is rougher than a dedicated implementation would be: the call still returns, it is just less polished:

• fr, it, eu, fa and other languages marked · for numbers_to_digits in the README matrix rely on the generic span-replacement fallback.
• Polish extraction does not merge tysiąc groups written with the singular form (jeden tysiąc jeden). Plural forms (dwa tysiące trzy) work.
• Czech pronounces four-digit numbers date-style (1234 → dvanáct třicet čtyři), which extraction does not reverse.

Full function parity

Every supported language provides every public function. Where a language has no hand-written implementation, a documented generic fallback fills in:

• pronounce_fraction reuses the language's nice_number fraction wording (which carries its plural/declension rules) and spells out any digits. The Slavic languages use feminine numerators (dvě třetiny, две трети) and Hungarian its két allomorph. Denominators outside the language's fraction vocabulary fall back to "cardinal numerator + ordinal denominator".
• is_ordinal falls back to a reverse lookup over pronounce_ordinal (1–100, round hundreds, 1000, 1000000), registering both -o/-a grammatical-gender variants.
• numbers_to_digits falls back to replacing maximal spoken-number spans found by extract_number, joining across connector words (vingt et un, sto in ena) only when the combined words read as one larger number. Languages whose hand-written converters stopped at twenty or mishandled compound and comma-grouped numbers (az, ca, cs, da, en, es, nl, pl) now use this path.

The parity guarantee is enforced by tests/test_lang_parity.py, which runs every public function against every supported language.

---

← API reference · Home · Adding a language →