hebrew-num2words

Turn digits in Hebrew text into spoken Hebrew words. Pure Python, no dependencies.

pip install hebrew-num2words

from hebrew_num2words import normalize_numbers

normalize_numbers("יש לי 3 ילדים")        # יש לי שלושה ילדים
normalize_numbers("יש לי 3 בנות")         # יש לי שלוש בנות
normalize_numbers("המחיר 1250 שקלים")     # המחיר אלף מאתיים וחמישים שקלים
normalize_numbers("נפגשים ב-15 במרץ")     # נפגשים בחמישה עשר במרץ
normalize_numbers("השעה 14:45")           # השעה רבע לשלוש
normalize_numbers("עלייה של 3.5%")        # עלייה של שלושה וחצי אחוז

Digit-free text is returned unchanged, byte for byte.

What it handles

• Gender agreement with the counted noun — 8 שעות → שמונה (feminine), 8 שקלים → שמונה (masculine); the noun after the numeral decides, which is the part naive expanders get wrong.
• Construct forms — 2 ספרים → שני, 2 מכוניות → שתי (not שניים / שתיים).
• Bare numerals — a count with no noun takes the absolute (feminine) series, except 1 → אחד.
• Clock times — 14:45 is spoken 12-hour and feminine, with ורבע / וחצי / רבע ל־ / עשרים ל־.
• Percents — 12% → masculine cardinal + אחוז.
• Dates — 3.5.2026 → masculine day + ב+month + feminine year.
• Identifiers — phone numbers, ID numbers and codes are read digit by digit.
• Ordinals and definite forms — ה-2 → השני.

API

normalize_numbers(text: str) -> str          # expand every numeric expression in a sentence
expand_token(token, prev_words=(), next_words=()) -> str
cardinal(n: int, gender: str, construct: bool = False) -> str    # 0..999,999
ordinal(n: int, gender: str, definite: bool = False) -> str

gender is MASC or FEM, exported from the package.

Where the rules come from

Every convention was mined from a 473-row Hebrew↔IPA gold set of spoken numbers, not hand-invented. On that set the expander takes word error rate from 50.04 to 13.52. 115 unit tests cover the cases above.

It exists because grapheme-to-phoneme models are typically trained on text that contains no digits at all, so a digit reaching the model produces guaranteed-wrong output — and often corrupts the word after it. Normalizing first removes the whole class.

License

Apache-2.0.