Language Support

Features

• Gold dictionary: 50k+ high-confidence entries

• Silver dictionary: Additional 50k+ entries

• POS-aware pronunciation: Different pronunciations based on part of speech

• Stress assignment: Primary and secondary stress markers

• Number handling: Cardinals, ordinals, currency

• Contraction support: Proper handling of “can’t”, “won’t”, etc.

Semantic ownership

English reviewed written-to-spoken preparation runs once per homogeneous language run through abbr2words and spokenform, covering shared abbreviations, dates, quantities, temperatures, currencies, and ordinary number forms. kokorog2p retains English typography, tokenization, alignment, and G2P behavior. Phoneme- sensitive number categories, including forms whose pronunciation depends on the English G2P context, remain downstream-owned by the local NumberConverter path.

Usage

from kokorog2p.en import EnglishG2P

# US English
g2p_us = EnglishG2P(
    language="en-us",
    use_espeak_fallback=True,
    use_spacy=True,
    spacy_model="en_core_web_md",  # explicit; unset selects the highest installed model
)

# British English
g2p_gb = EnglishG2P(
    language="en-gb",
    use_espeak_fallback=True,
    use_spacy=True,
    spacy_model="en_core_web_md",  # explicit; unset selects the highest installed model
)

# Optional: automatic selection or an exact model
g2p_auto = EnglishG2P(language="en-us", use_spacy=True)
g2p_sm = EnglishG2P(language="en-us", use_spacy=True, spacy_model="en_core_web_sm")

Examples

from kokorog2p import phonemize

# Context-dependent pronunciation
print(phonemize("I read a book.", language="en-us"))
# → ˈaɪ ɹˈɛd ə bˈʊk.

print(phonemize("I will read tomorrow.", language="en-us"))
# → ˈaɪ wɪl ɹˈid təmˈɑɹO.

# Numbers and currency
print(phonemize("I paid $1,234.56 for it.", language="en-us"))
# → aɪ pˈeɪd wʌn θˈaʊzənd tˈu hˈʌndɹəd...

Features

• Large dictionary: 738k+ entries with stress markers

• Phonological rules:

  • Final obstruent devoicing (Auslautverhärtung)

  • ich-Laut [ç] vs ach-Laut [x] alternation

  • Word-initial sp/st → [ʃp]/[ʃt]

  • Vowel length rules

  • Schwa in unstressed syllables

• Deterministic normalization: German cardinals, decimals, thousands groups, years, conservative ordinals, dates, times, EUR amounts, temperatures, and numbered units are classified before lexical abbreviation expansion.

• Abbreviations and units: Flexible z.B., d.h., and u.a. forms, common lexical abbreviations, and grammatical unit forms such as 1 Std. → eine Stunde and 2 kg → zwei Kilogramm.

• Ambiguity policy: Bare sentence-final numbers stay cardinals, invalid dates and times remain unchanged, and unit symbols only expand when preceded by a number. The exact written-to-spoken result, including abbreviation and initialism choices, is supplied by the supported bounded Spokenform 0.2.x profile rather than a second kokorog2p rule set.

• Regional variants: de-de, de-at, de-ch

German was the first language migrated to the shared semantic-preparation architecture, followed by French, Spanish, Italian, Portuguese, Czech, and English. These seven languages use the same per-homogeneous-run spokenform path. abbr2words remains the shared source of truth for lexical abbreviation and symbol recognition, while spokenform owns accepted written-to-spoken behavior through its for_kokorog2p profile. The examples in this guide are illustrative; the supported bounded Spokenform 0.2.x adapter API is authoritative. Kokorog2p retains only typography and phoneme-sensitive handling for spans Spokenform leaves protected or unsupported.

Usage

from kokorog2p.de import GermanG2P

g2p = GermanG2P(
    language="de-de",
    use_espeak_fallback=True,
    strip_stress=True
)

Examples

from kokorog2p import phonemize

# Basic phonemization
print(str(phonemize("Guten Tag", language="de")))
# → ɡuːtn̩ taːk

# Phonological rules
print(str(phonemize("ich", language="de")))      # → ɪç (ich-Laut)
print(str(phonemize("ach", language="de")))      # → ax (ach-Laut)
print(str(phonemize("Tag", language="de")))      # → taːk (final devoicing)

# Numbers
print(str(phonemize("Ich habe 42 Euro.", language="de")))
# → ɪç haːbə t͡svaɪ̯ʊntfɪɐ̯t͡sɪç ɔɪ̯ʁo.

Features

• Gold dictionary: High-quality French pronunciations

• Semantic ownership: abbr2words recognizes French abbreviations and symbols; spokenform prepares dates, times, numbers, ordinals, currencies, quantities, units, and temperatures

• espeak-ng fallback: For out-of-vocabulary words

French was the second migrated language. kokorog2p retains French typography, tokenization, lexicon lookup, fallback, and phoneme conversion. The legacy helpers in kokorog2p.fr.numbers remain as deprecated compatibility wrappers.

Usage

from kokorog2p.fr import FrenchG2P

g2p = FrenchG2P(
    language="fr-fr",
    use_espeak_fallback=True
)

Examples

from kokorog2p import phonemize

print(phonemize("Bonjour le monde", language="fr"))
# → bɔ̃ʒuʁ lə mɔ̃d

print(phonemize("J'ai vingt et un ans.", language="fr"))
# → ʒɛ vɛ̃t e œ̃ ɑ̃.

Features

• Rule-based phonology:

  • Palatalization (d+i → ɟ, t+i → c, n+i → ɲ)

  • Long vowels (á → aː, í → iː, etc.)

  • ř phoneme [r̝]

  • ch digraph → [x]

  • Final devoicing

  • Voicing assimilation

• No dictionary required: Works with any Czech text

• Semantic ownership: abbr2words owns Czech abbreviation and canonical quantity/currency recognition; spokenform owns reusable numbers, dates, quantities, temperatures, and currencies; kokorog2p retains Czech typography, tokenization, alignment, phonology, lexicon, and fallbacks.

Czech semantic preparation runs once per homogeneous language run. The Czech normalizer is a downstream typography adapter, so direct CzechG2P use and the public pipeline share the same spokenform preparation. Czech times remain caller-managed.

Usage

from kokorog2p.cs import CzechG2P

g2p = CzechG2P(language="cs-cz")

Examples

from kokorog2p import phonemize

print(phonemize("Dobrý den", language="cs"))
# → dobriː dɛn

print(phonemize("Praha", language="cs"))
# → praɦa

# Palatalization
print(phonemize("děti", language="cs"))
# → ɟɛcɪ

 # ř phoneme
 print(phonemize("řeka", language="cs"))
 # → r̝ɛka

Features

• Rule-based phonology:

  • 5 pure vowels (a, e, i, o, u)

  • Stress prediction (penultimate for vowel-ending, final for consonant-ending)

  • Palatal sounds: ñ [ɲ], ll [ʎ] or [j]

  • Jota: j/g+e/i [x]

  • Theta: z/c+e/i [θ] (European) or [s] (Latin American)

  • Tap vs trill: r [ɾ] vs rr [r]

• Dialect support: es (European), la (Latin American)

• Number handling: Cardinals, ordinals, currency

• Semantic ownership: abbr2words recognizes Spanish abbreviations and symbols; spokenform prepares written numbers, quantities, temperatures, currencies, and reviewed dates; kokorog2p retains typography, tokenization, alignment, and G2P

Spanish semantic preparation is applied once per homogeneous language run. Dialect selection remains a phoneme-layer concern: es provides European theta behavior and la provides Latin-American seseo behavior over the same prepared text.

Usage

from kokorog2p.es import SpanishG2P

g2p = SpanishG2P(
    language="es",
    dialect="es"  # or "la" for Latin American
)

Examples

from kokorog2p import phonemize

print(phonemize("Hola mundo", language="es"))
# → ola mundo

# Phonological features
print(phonemize("año", language="es"))      # → aɲo
print(phonemize("calle", language="es"))    # → kaʎe or kaje
print(phonemize("perro", language="es"))    # → pero (trilled r)

Semantic ownership

• abbr2words owns Italian lexical abbreviations, units, currency symbols, canonical structured IDs, and source-span matching.

• spokenform owns reviewed dates, quantities, temperatures, currencies, and ordinary written numbers for each homogeneous Italian run.

• kokorog2p retains Italian typography, apostrophe/contraction handling, tokenization, spaCy integration, and G2P stress, gemination, phoneme, and vocabulary behavior.

Italian colon times remain caller-managed in this migration. spokenform does not perform language detection, mixed-language segmentation, markup parsing, or phoneme generation.

Features

• Rule-based phonology:

  • 5 pure vowels (a, e, i, o, u) - no reduction

  • Predictable stress (usually penultimate)

  • Gemination (double consonants) preservation

  • Palatals: gn [ɲ], gli [ʎ]

  • Affricates: z [ʦ/ʣ], c/ci [ʧ], g/gi [ʤ]

  • Context-sensitive c/g pronunciation

• Stress marking: Automatic stress detection from accents

• Number handling: Cardinals, ordinals

Usage

from kokorog2p.it import ItalianG2P

g2p = ItalianG2P(
    language="it-it",
    mark_stress=True,
    mark_gemination=True
)

Examples

from kokorog2p import phonemize

print(phonemize("Ciao mondo", language="it"))
# → ʧao mondo

# Gemination
print(phonemize("anno", language="it"))     # → anːo
print(phonemize("fatto", language="it"))    # → fatːo

# Palatals
print(phonemize("gnocchi", language="it"))  # → ɲɔkːi
print(phonemize("figlio", language="it"))   # → fiʎo

Features

• Rule-based phonology:

  • 7 oral vowels (a, e, ɛ, i, o, ɔ, u)

  • 5 nasal vowels (ã, ẽ, ĩ, õ, ũ)

  • Nasal diphthongs

  • Palatalization: lh [ʎ], nh [ɲ], x/ch [ʃ]

  • Affrication: t+i [ʧ], d+i [ʤ] (Brazilian)

  • Sibilants: s [s/z], x [ʃ], z [z]

  • Liquids: r [ʁ/x/h], rr [ʁ/x], single r [ɾ]

• Dialect: Brazilian Portuguese (pt-br) by default; semantic preparation also supports European Portuguese wording (pt-pt)

• Stress marking: Automatic stress assignment

Usage

from kokorog2p.pt import PortugueseG2P

g2p = PortugueseG2P(
    language="pt-br",
    mark_stress=True,
    affricate_ti_di=True  # Brazilian feature
)

Examples

from kokorog2p import phonemize

print(phonemize("Olá mundo", language="pt"))
# → ola mundo

# Nasal vowels
print(phonemize("mãe", language="pt"))      # → mãj̃
print(phonemize("pão", language="pt"))      # → pãw̃

# Affrication (Brazilian)
print(phonemize("tia", language="pt"))      # → ʧia
print(phonemize("dia", language="pt"))      # → ʤia

Features

• Jieba tokenization: Chinese word segmentation

• Pypinyin conversion: Pinyin to IPA

• Tone sandhi: Automatic tone changes

• cn2an: Number to Chinese conversion

• Punctuation mapping: Chinese to Western punctuation

Usage

from kokorog2p.zh import ChineseG2P

g2p = ChineseG2P(
    language="zh",
    version="1.1"
)

Examples

from kokorog2p import phonemize

print(phonemize("你好世界", language="zh"))
# → nǐ hǎo shì jiè (with tone markers)

Features

• pyopenjtalk: Full Japanese text analysis

• Mora-based: Phonemes aligned with mora structure

• Pitch accent: Automatic pitch accent assignment

• Number handling: Japanese numerals

Usage

from kokorog2p.ja import JapaneseG2P

g2p = JapaneseG2P(
    language="ja",
    version="pyopenjtalk"
)

Examples

from kokorog2p import phonemize

print(phonemize("こんにちは", language="ja"))
# → koɴɲit͡ɕiha

print(phonemize("世界", language="ja"))
# → sekai

Features

• MeCab integration: Korean morphological analysis

• Phonological rules:

  • Consonant assimilation

  • Palatalization

  • Tensification

  • Aspiration

  • Liaison (연음)

  • Final consonant neutralization

• Hanja support: Sino-Korean character handling

• Number handling: Korean numerals

Usage

from kokorog2p.ko import KoreanG2P

g2p = KoreanG2P(
    language="ko-kr",
    use_mecab=True
)

Examples

from kokorog2p import phonemize

print(phonemize("안녕하세요", language="ko"))
# → annjʌŋhasejo

# Phonological rules
print(phonemize("학교", language="ko"))     # → hakk͈jo (tensification)
print(phonemize("받침", language="ko"))     # → patʃʰim (palatalization)

Features

• phonikud integration: Hebrew nikud to IPA conversion

• Nikud handling: Processes diacritical marks for vowels

• Stress prediction: Automatic stress assignment

• Modern Hebrew: Optimized for contemporary pronunciation

Usage

from kokorog2p.he import HebrewG2P

g2p = HebrewG2P(
    language="he-il",
    preserve_punctuation=True,
    preserve_stress=True
)

Examples

from kokorog2p import phonemize

# Requires nikud (diacritical marks)
print(phonemize("שָׁלוֹם", language="he"))
# → ʃalom

print(phonemize("עִבְרִית", language="he"))
# → ivʁit

Features

• Automatic detection: Word-level language detection using lingua-py

• High accuracy: >90% accuracy for words with 5+ characters

• Caching: Detection results cached for performance

• Configurable threshold: Control detection sensitivity

• Graceful degradation: Falls back to primary language without lingua-py

• 17+ languages: Support for major world languages

Supported Languages

• English (en-us, en-gb)

• German (de)

• French (fr)

• Spanish (es)

• Italian (it)

• Portuguese (pt)

• Japanese (ja)

• Chinese (zh)

• Korean (ko)

• Hebrew (he)

• Czech (cs)

• Dutch (nl)

• Polish (pl)

• Russian (ru)

• Arabic (ar)

• Hindi (hi)

• Turkish (tr)

Usage

from kokorog2p import phonemize
from kokorog2p.multilang import preprocess_multilang

text = "Das Meeting war great!"
overrides = preprocess_multilang(
    text,
    default_language="de",
    allowed_languages=["de", "en-us"],
)

result = phonemize(text, language="de", overrides=overrides)

Examples

German with English:

from kokorog2p import phonemize
from kokorog2p.multilang import preprocess_multilang

text = "Ich gehe zum Meeting. Let's discuss the Roadmap!"
overrides = preprocess_multilang(
    text,
    default_language="de",
    allowed_languages=["de", "en-us"],
)
result = phonemize(text, language="de", overrides=overrides)
print(result.phonemes)

English with German:

overrides = preprocess_multilang(
    "Hello, mein Freund! This is wunderbar.",
    default_language="en-us",
    allowed_languages=["en-us", "de"],
)
result = phonemize(
    "Hello, mein Freund! This is wunderbar.",
    language="en-us",
    overrides=overrides,
)
print(result.phonemes)

Multiple languages:

overrides = preprocess_multilang(
    "Bonjour! The Meeting ist wichtig.",
    default_language="fr",
    allowed_languages=["fr", "en-us", "de"],
)
result = phonemize(
    "Bonjour! The Meeting ist wichtig.",
    language="fr",
    overrides=overrides,
)
print(result.phonemes)

Configuration

Confidence threshold:

from kokorog2p.multilang import preprocess_multilang

# Conservative (higher confidence required)
overrides = preprocess_multilang(
    "Das Meeting ist wichtig",
    default_language="de",
    allowed_languages=["de", "en-us"],
    confidence_threshold=0.9,  # Default: 0.7
)

# Aggressive (lower confidence required)
overrides = preprocess_multilang(
    "Das Meeting ist wichtig",
    default_language="de",
    allowed_languages=["de", "en-us"],
    confidence_threshold=0.5,
)

How It Works

1. Text is tokenized into words

2. Each word is sent to the language detector

3. Detector returns language + confidence score

4. If confidence ≥ threshold and language is allowed:

   • An OverrideSpan is created with {"lang": "..."}

   • Short words (<3 chars) keep the default language

Performance

• Memory: ~100 MB for lingua models (loaded once)

• Speed: ~0.1-0.5 ms per word

• Accuracy: >90% for words with 5+ characters

Limitations

• Short words (<3 characters) use the default language only

• Proper nouns may be misdetected

• Requires lingua-language-detector installation

• Detection quality varies by word distinctiveness

Installation

pip install kokorog2p[mixed]

English

from kokorog2p.en.numbers import expand_number

print(expand_number("I have $42.50"))
# → I have forty-two dollars and fifty cents

German

from kokorog2p.de.numbers import expand_number

print(expand_number("Ich habe 42 Euro."))
# → Ich habe zweiundvierzig Euro.

French

from kokorog2p.fr.numbers import expand_number

print(expand_number("J'ai 42 euros."))
# → J'ai quarante-deux euros.