Bulgarian Text Normalizer for TTS

A comprehensive text normalization package that converts written Bulgarian text into its spoken form, designed as a preprocessing step for Text-to-Speech (TTS) systems.

Features

Grammatical Correctness

• Gender agreement: Handles masculine/feminine/neuter (един/една/едно, два/две)
• Ordinal forms: Full gender-aware ordinals (първи/първа/първо)
• Year reading: Ordinal feminine form matching "година" (две хиляди двадесет и шеста)
• Space-separated thousands: 7 000 000 → седем милиона

Usage

Quick usage

from bg_text_normalizer import normalize_text

result = normalize_text("На 15.02.2026 г. в 14:30 ч. цената е 1500.50 лв.")
# "На петнадесети февруари две хиляди двадесет и шеста година в четиринадесет
#  и тридесет часа цената е хиляда и петстотин лева и петдесет стотинки."

Class-based usage

from bg_text_normalizer import BulgarianTextNormalizer

normalizer = BulgarianTextNormalizer(expand_abbrevs=True, verbose=False)
result = normalizer.normalize("бул. Витоша №10, гр. София")
# "булевард Витоша номер десет, град София"

A Roman numeral after a capitalized name (Карл V) is left alone by default, since Марк I can be a product name. Turn it on with roman_numerals=True:

BulgarianTextNormalizer(roman_numerals=True).normalize("Борис III управлява.")
# "Борис трети управлява."

A lone C, D, L or M after a name is skipped, because it is usually an initial.

Individual modules

from bg_text_normalizer.bg_numbers import number_to_words_cardinal, number_to_words_ordinal
from bg_text_normalizer.bg_dates import normalize_date
from bg_text_normalizer.bg_currency import normalize_currency

number_to_words_cardinal(2500, gender='m')    # "две хиляди и петстотин"
number_to_words_ordinal(15, gender='m')       # "петнадесети"
normalize_date(15, 2, 2026)                   # "петнадесети февруари две хиляди двадесет и шеста"
normalize_currency("99.99", "BGN")            # "деветдесет и девет лева и деветдесет и девет стотинки"

Integration with TTS Training (Qwen3-TTS)

Use this normalizer as a preprocessing step when preparing your training data:

import json
from bg_text_normalizer import normalize_text

# Process your JSONL training data
with open('raw_data.jsonl', 'r') as f_in, open('normalized_data.jsonl', 'w') as f_out:
    for line in f_in:
        entry = json.loads(line)
        entry['text'] = normalize_text(entry['text'])
        f_out.write(json.dumps(entry, ensure_ascii=False) + '\n')

For inference (runtime TTS), add normalization before synthesis:

from bg_text_normalizer import normalize_text

def synthesize(text: str):
    normalized = normalize_text(text)
    # ... pass normalized text to TTS model

File Structure

bg-text-normalizer/
├── src/
│   └── bg_text_normalizer/
│       ├── __init__.py           # Package entry point
│       ├── bg_normalizer.py      # Main orchestrator
│       ├── bg_numbers.py         # Cardinal, ordinal, decimal numbers
│       ├── bg_dates.py           # Date normalization
│       ├── bg_time.py            # Time normalization
│       ├── bg_currency.py        # Currency (BGN, EUR, USD, GBP)
│       ├── bg_scales.py          # хил., млн., млрд.
│       ├── bg_units.py           # Measurement units with count forms
│       ├── bg_punctuation.py     # Keeps an abbreviation's period only at sentence end
│       ├── bg_abbreviations.py   # 100+ Bulgarian abbreviations
│       ├── bg_phone.py           # Phone number reading
│       └── bg_roman.py           # Roman numeral conversion
├── test_normalizer.py            # Test suite
├── pyproject.toml
└── README.md

Adding Custom Abbreviations

Edit src/bg_text_normalizer/bg_abbreviations.py and add entries to the appropriate dictionary:

# In ADDRESS_ABBREVS, TITLE_ABBREVS, etc.
CUSTOM_ABBREVS = {
    'your_abbrev.': 'пълна форма',
}

Dependencies

None — pure Python, no external dependencies required.