syllable-counter 0.1.0

Multi-language syllable counter and TTS duration estimator (12 languages)

syllable-counter

Multi-language syllable counter and TTS duration estimator. Supports 12 languages with per-language strategies (pyphen hyphenation, pypinyin, regex on Hangul blocks, optional pyopenjtalk/phonikud).

Install

pip install syllable-counter

Optional native backends for better accuracy:

pip install syllable-counter[ja]   # adds pyopenjtalk for Japanese
pip install syllable-counter[he]   # adds phonikud for Hebrew
pip install syllable-counter[all]  # both

Usage

from syllable_counter import count_syllables, estimate_duration

# Count syllables
count_syllables("hello world", lang="en")       # 3
count_syllables("xin chào các bạn", lang="vi")  # 4
count_syllables("我是中国人", lang="zh")          # 5
count_syllables("iPhone15", lang="en")           # iPhone + 15 = ~6

# Numbers are read in the target language
count_syllables("2026", lang="en")  # "two thousand and twenty-six"  → 7
count_syllables("2026", lang="es")  # "dos mil veintiséis"           → 4
count_syllables("2026", lang="zh")  # "二零二六"                       → 4

# Estimate TTS duration
est = estimate_duration("hello world", lang="en")
# DurationEstimate(seconds=0.87, frames=47, syllable_count=3, ...)

Supported languages

Code	Language	Backend
en	English	pyphen + vowel-group fallback
es	Spanish	pyphen + vowel-group fallback
fr	French	pyphen + vowel-group fallback
de	German	pyphen + vowel-group fallback
it	Italian	pyphen + vowel-group fallback
pt	Portuguese	pyphen + vowel-group fallback
ru	Russian	pyphen + per-vowel Cyrillic fallback
vi	Vietnamese	whitespace tokens (monosyllabic)
zh	Chinese	pypinyin
ja	Japanese	pyopenjtalk (optional) → char-range fallback
ko	Korean	regex on Hangul blocks
he	Hebrew	phonikud (optional) → CV-structure fallback

Numbers are handled per-language via num2words (45 languages natively, with English fallback for the rest, plus digit-by-digit native fallback for Chinese).

Calibrated duration estimation

When you have reference word timestamps from a target speaker, you can calibrate:

ref_words = [
    ("hello", 0.0, 0.4),
    ("world", 0.5, 0.9),
    ("how", 1.0, 1.2),
    ("are", 1.3, 1.5),
    ("you", 1.6, 1.8),
]
est = estimate_duration("nice to meet you", lang="en", ref_words=ref_words)
est.calibrated           # True
est.sec_per_syllable_used  # fitted from the reference

Calibration uses OLS to fit word_duration ≈ silence + sec_per_syl × n_syllables. Falls back to per-language defaults when fewer than 5 usable reference words are provided.

Pause-aware

count_pauses(text) returns (short, medium, long) punctuation counts:

count_pauses("hello, world. yes!")  # (1, 0, 2)

estimate_duration automatically adds pause time:

• short (,, 、, ，) = 1× short_pause
• medium (;, :, ；, ：) = 1.5× short_pause
• long (., !, ?, 。, ！, ？) = 2.5× short_pause

API

count_syllables(text: str, lang: str) -> int
compute_sec_per_syllable(ref_words, lang) -> (sec_per_syl, silence_overhead)
compute_short_pause(ref_words) -> float
count_pauses(text: str) -> (short, medium, long)
estimate_duration(target_text, lang, ref_words=None, ...) -> DurationEstimate

Errors:

• UnsupportedLanguageError — lang not in SUPPORTED_LANGS
• InsufficientReferenceError — calibration needs ≥5 usable words

Limitations

• Pyphen-based fallback for unknown words is English-tuned for silent-e
• '5.5' reads as "five five" (decimal point is lost; off by 1 syl)
• '-5' reads as just "five" (negative sign lost; off by 2 syl)
• Hebrew without nikud uses ceil(letters/2) approximation (±1 syl per word)
• Japanese without pyopenjtalk treats each kanji as 1 mora (kanji can be 2–3)

License

MIT