canto-hk-g2p

Fast Cantonese text-to-phoneme (G2P) converter — converts Cantonese (and English code-switched) text to Jyutping with tone numbers (LSHK standard, all-ASCII) or IPA. Rust core with Python bindings via PyO3/maturin.

from canto_hk_g2p import Pipeline

p = Pipeline()
p.convert("你好嘅，I love Hong Kong")
# → "nei5 hou2 ge3 ， I love Hong Kong"

Why this exists

Cantonese TTS systems struggle with two fundamental problems. First, colloquial Cantonese characters (嘅, 喺, 噉, 㗎, …) are rare or absent from standard Chinese vocabulary tables, causing tokenizers to split them into byte fragments that the model never learns to pronounce reliably. Second, Hong Kong Cantonese freely mixes English — a sentence like 佢 send 咗 email 俾我 is everyday speech — and no existing Cantonese G2P tool handles this cleanly.

canto-hk-g2p solves both: it converts the entire input, including colloquial particles, numbers, dates, and inline English, to all-ASCII Jyutping before the text ever reaches a TTS encoder. The result is stable, tone-accurate phoneme sequences with 100% coverage.

The library is a standalone open-source deliverable — Cantonese TTS pre-processing is the primary motivation, but the tool is useful anywhere Cantonese phonemization is needed.

Features

• All-ASCII Jyutping output with LSHK tone numbers (nei5 hou2 ge3)
• IPA output — convert_ipa() with tone diacritics (nei̯˩˧ hou̯˧˥ kɛː˧) or tone numbers (nei̯5 hou̯2 kɛː3); English tokens converted via CMU Pronouncing Dictionary
• English code-switching — the only Cantonese G2P tool that handles mixed HK text; English tokens pass through unchanged in Jyutping mode (佢 send 咗 email 俾我 → keoi5 send zo2 email bei2 ngo5), or converted to IPA in IPA mode
• Punctuation normalisation (punc_norm=True by default) — converts exotic punctuation to TTS-friendly equivalents before G2P: 「」《》 removed, … → 。, —— → ，, · → space, 、 → ，
• Text normalization — Arabic and full-width digits expanded to Cantonese spoken form:
  • Year: 2026年 → ji6 ling4 ji6 luk6 nin4
  • Date: 6月13日 → luk6 jyut6 sap6 saam1 jat6
  • Percentage: 50% → baak3 fan6 zi1 ng5 sap6; decimal 50.5% → baak3 fan6 zi1 ng5 sap6 dim2 ng5
  • Currency symbols: HK$100 → jat1 baak3 jyun4; ¥500 → ng5 baak3 jat6 jyun4; €200 → ji6 baak3 au1 jyun4
  • Currency codes: USD100 → jat1 baak3 mei5 jyun4; EUR 200 → ji6 baak3 au1 jyun4
  • Measurement units: 120km/h → 一百二十公里每小時; 36.5°C → 三十六點五攝氏度; 75kg → 七十五公斤
  • Fractions: 1/2 → 二分之一; 3/4 → 四分之三
  • Scores: 3:1 → 三比一; 10:0 → 十比零
  • Ordinal/floor/episode: 第3名 → 第三名; 3樓 → 三樓; 第3集 → 第三集
  • Time: 下午3時15分 → Cantonese spoken time
  • Phone numbers: digit-by-digit expansion
• Polyphone disambiguation via longest-match word-level segmentation (~85% accuracy)
• user_dict runtime overrides — Pipeline(user_dict={"行": "hong4"}) locks a pronunciation above every bundled dictionary, and participates in segmentation so multi-char overrides aren't split apart
• convert_candidates() / convert_candidates_batch() — Candidates API: rank-ordered alternate readings for polyphones (正經 → ["zing3 ging1", "zing1 ging1"]), instead of committing to a single one
• Confidence + source tags — every structured result carries a categorical confidence tag ("certain" / "ranked" / "tied") and a data-layer source tag ("rime" / "tojyutping" / "oral_hk" / "variant_alias" / "unihan" / "user_dict" / ...) per token
• 借音字 (phonetic-loan) correction — common miswritings that borrow a homophone's sound (訓覺 for 瞓覺, 岩岩 for 啱啱) resolve to the correctly-spelled word's reading instead of the variant character's own native reading (data/variant_words.tsv)
• Rayon parallel batch processing — scales to large corpora
• Zero runtime Python dependencies — pronunciation dictionaries bundled in the wheel
• convert_detailed() / convert_detailed_batch() — structured (token, jyutping, lang, confidence, source) output
• Apache-2.0 license, permissive data sources only — safe to redistribute

Installation

# pip
pip install canto-hk-g2p

# uv (recommended — faster resolver)
uv pip install canto-hk-g2p

# uv project
uv add canto-hk-g2p

Pre-built wheels are available for Linux (x86_64, aarch64), macOS (x86_64, Apple Silicon), and Windows (x86_64) for Python 3.8+. No Rust toolchain required for end users.

Quick start

from canto_hk_g2p import Pipeline

p = Pipeline()

# Basic Cantonese
p.convert("你好嘅")
# → "nei5 hou2 ge3"

# Full sentence with punctuation and English
p.convert("你好嘅，I love Hong Kong")
# → "nei5 hou2 ge3 ， I love Hong Kong"

# Punctuation normalisation (default on)
p.convert("《天氣之子》——一個故事")
# → "tin1 hei3 zi1 zi2 ， jat1 go3 gu3 si6"

# Number and date normalization
p.convert("2026年6月13日")
# → "ji6 ling4 ji6 luk6 nin4 luk6 jyut6 sap6 saam1 jat6"

p.convert("50%")
# → "baak3 fan6 zi1 ng5 sap6"

# English code-switching — Latin tokens pass through unchanged
p.convert("佢send咗email俾我")
# → "keoi5 send zo2 email bei2 ngo5"

# Batch conversion (Rayon parallel)
p.convert_batch(["香港", "銀行", "你好嘅"])
# → ["hoeng1 gong2", "ngan4 hong4", "nei5 hou2 ge3"]

# Structured output with language, confidence, and source tags
p.convert_detailed("香港 hello")
# → [("香港", "hoeng1 gong2", "yue", "certain", "rime"),
#     ("hello", "hello", "en", "certain", "passthrough")]

IPA output

from canto_hk_g2p import Pipeline, jyutping_to_ipa

p = Pipeline()

# Cantonese → IPA with tone diacritics (default)
p.convert_ipa("你好嘅")
# → "nei̯˩˧ hou̯˧˥ kɛː˧"

# IPA with tone numbers
p.convert_ipa("你好嘅", tone="number")
# → "nei̯5 hou̯2 kɛː3"

# English code-switching → English tokens converted via CMU Pronouncing Dictionary
p.convert_ipa("佢 send 咗 email 俾我")
# → "kʰœːy̯˩˧ sɛnd tsɔː˧˥ iːmeɪl pei̯˧˥ ŋɔː˩˧"

# Unknown English words (OOV) pass through unchanged
p.convert_ipa("佢去MTR站")
# → "kʰœːy̯˩˧ hœːy̯˧˥ MTR tsaam6˨"

# Standalone utility — convert existing Jyutping strings to IPA
jyutping_to_ipa("hoeng1 gong2")
# → "hœːŋ˥ kɔːŋ˧˥"

jyutping_to_ipa("nei5 hou2 ge3", tone="number")
# → "nei̯5 hou̯2 kɛː3"

IPA tone marks: ˥ high level (1), ˧˥ high rising (2), ˧ mid level (3), ˨˩ low falling (4), ˩˧ low rising (5), ˨ low level (6).

jyutping_to_ipa is also importable directly from canto_hk_g2p as shown above, or from canto_hk_g2p.ipa.

API reference

Pipeline(*, punc_norm=True, user_dict=None)

Loads the binary pronunciation dictionaries from the bundled data/ directory. The pipeline object is reusable and thread-safe for batch calls.

p = Pipeline()                   # punc_norm on (default)
p2 = Pipeline(punc_norm=False)   # raw punctuation passthrough

user_dict — runtime pronunciation override

p3 = Pipeline(user_dict={"行為": "hang4 wai4", "老世": "lou5 sai3"})
p3.convert("行為")   # → "hang4 wai4"   (overrides the bundled dict)
p3.convert("老世")   # → "lou5 sai3"    (not in any bundled dict at all)

Each value must have exactly as many space-separated syllables as the key has characters, and every syllable must be valid Jyutping:

Pipeline(user_dict={"老世": "lou5"})     # ValueError — 2 chars, 1 syllable
Pipeline(user_dict={"行": "xyz9"})       # ValueError — not a valid syllable

Known limitation: an override only competes with the bundled word_dict at its own starting position during segmentation. If a longer dictionary word starts one character earlier and contains the override's key as a substring, that longer word wins and the override is shadowed:

p = Pipeline(user_dict={"正經": "zing1 ging1"})
p.convert("正經嚟講")   # → "zing1 ging1 lai4 gong2"  (override applied)
p.convert("佢好正經")   # → "keoi5 hou2 zing3 ging1"  (shadowed — "好正經" is itself
                        #    a 3-char dictionary entry starting one char earlier)

convert(text: str) -> str

Converts a string of Cantonese (or mixed Cantonese/English) text to space-separated Jyutping syllables.

Returns: str — space-separated Jyutping syllables. Punctuation tokens are preserved as-is. Latin tokens (English words, acronyms) are passed through unchanged.

p.convert("銀行")          # → "ngan4 hong4"
p.convert("2026年")        # → "ji6 ling4 ji6 luk6 nin4"
p.convert("I love you")   # → "I love you"
p.convert("send")          # → "send"
p.convert("")              # → ""

convert_batch(texts: list[str]) -> list[str]

Converts a list of strings in parallel using Rayon. Output order matches input order.

Returns: list[str] — one Jyutping string per input, in the same order.

p.convert_batch(["香港", "銀行", "你好嘅"])
# → ["hoeng1 gong2", "ngan4 hong4", "nei5 hou2 ge3"]

p.convert_batch([])   # → []

convert_ipa(text: str, tone: str = "diacritic") -> str

Converts text to IPA. Cantonese tokens use the Jyutping→IPA mapping table; English tokens are converted via the bundled CMU Pronouncing Dictionary (OOV words pass through unchanged).

Returns: str — space-separated IPA string.

p.convert_ipa("香港")
# → "hœːŋ˥ kɔːŋ˧˥"

p.convert_ipa("香港", tone="number")
# → "hœːŋ1 kɔːŋ2"

p.convert_ipa("send")     # English → CMU dict
# → "sɛnd"

p.convert_ipa("MTR")      # OOV → passthrough
# → "MTR"

jyutping_to_ipa(jyutping: str, tone: str = "diacritic") -> str

Standalone utility — converts an existing space-separated Jyutping string to IPA. Non-Jyutping tokens (English, punctuation) pass through unchanged. Importable from canto_hk_g2p or canto_hk_g2p.ipa.

from canto_hk_g2p import jyutping_to_ipa

jyutping_to_ipa("nei5 hou2 ge3")
# → "nei̯˩˧ hou̯˧˥ kɛː˧"

jyutping_to_ipa("hoeng1 gong2", tone="number")
# → "hœːŋ1 kɔːŋ2"

confidence and source — shared across convert_detailed() / convert_candidates()

Every structured result (convert_detailed(), convert_candidates(), and their batch siblings) carries two trailing fields:

No numeric probability is exposed, by design. Neither ToJyutping's trie nor rime-cantonese's tied readings carry real frequency data — a float score per candidate would be fabricated, not measured. See CHANGELOG (v2.0.0) for the research behind this categorical-only design.

convert_detailed(text: str) -> list[tuple[str, str, str, str, str]]

Returns a structured token-level breakdown. Each element is a 5-tuple (token, jyutping, lang, confidence, source).

The joined jyutping from convert_detailed() is identical to the output of convert() for the same input.

p.convert_detailed("香港 hello")
# → [("香港", "hoeng1 gong2", "yue", "certain", "rime"),
#     ("hello", "hello", "en", "certain", "passthrough")]

p.convert_detailed("你好，")
# → [("你", "nei5", "yue", "ranked", "tojyutping"), ("好", "hou2", "yue", "ranked", "tojyutping"),
#     ("，", "，", "punct", "certain", "passthrough")]

p.convert_detailed("2026年")
# → all tokens tagged "yue" (normalizer expands digits to Chinese characters first)

p.convert_detailed("")
# → []

convert_detailed_batch(texts: list[str]) -> list[list[tuple[str, str, str, str, str]]]

Rayon-parallel batch sibling of convert_detailed() — same per-text output shape, one list per input text.

convert_candidates(text: str) -> list[tuple[str, list[str], str, str, str]]

The Candidates API — the text-level sibling of convert_detailed(). Returns (token, candidate_readings, lang, confidence, source) tuples, where candidate_readings is a rank-ordered list (most-likely first) instead of a single committed string.

candidate_readings[0] always equals the reading convert()/convert_detailed() commits to for that token. A list of length 1 means the bundled data has no known alternate reading for that exact token (this is the common case — most words and characters are unambiguous). Ambiguity is only reported at the granularity the bundled dictionaries track it: a whole multi-char word, or a single character. A user_dict override (see above) always collapses to a single candidate — an override is a final decision, not ambiguity to report.

p.convert_candidates("正經")
# → [("正經", ["zing3 ging1", "zing1 ging1"], "yue", "tied", "tojyutping_tiebreak")]

p.convert_candidates("香港")
# → [("香港", ["hoeng1 gong2"], "yue", "certain", "rime")]   (no known ambiguity)

p.convert_candidates("行")
# → [("行", ["haang4", "hang4", "hong4", ...], "yue", "ranked", "tojyutping")]

p2 = Pipeline(user_dict={"正經": "zing1 ging1"})
p2.convert_candidates("正經")
# → [("正經", ["zing1 ging1"], "yue", "certain", "user_dict")]   (override collapses ambiguity)

Known limitation: ambiguity is not surfaced across an out-of-vocabulary multi-char token's individual characters — such a token falls back to convert_detailed()'s single resolved reading (tagged "char_fallback"). In practice this case cannot actually be reached through convert_candidates(): the segmenter only ever emits a multi-char token when it's an exact word_dict/user_dict hit, so every multi-char token already has a definite reading by construction.

convert_candidates_batch(texts: list[str]) -> list[list[tuple[str, list[str], str, str, str]]]

Rayon-parallel batch sibling of convert_candidates() — same per-text output shape, one list per input text:

p.convert_candidates_batch(["正經", "香港銀行"])
# → [
#      [("正經", ["zing3 ging1", "zing1 ging1"], "yue", "tied", "tojyutping_tiebreak")],
#      [("香港", ["hoeng1 gong2"], "yue", "certain", "rime"),
#       ("銀行", ["ngan4 hong4"], "yue", "certain", "tojyutping")],
#    ]

Accuracy

Verified examples

Known limitations

• Residual polyphones: Word-boundary segmentation resolves approximately 85% of polyphone cases. Single-character polyphones that cannot be disambiguated by context (e.g. 好 as hou2 greeting vs. hou3 adverb) fall back to the most-frequent reading in the dictionary.
• 借音字 (phonetic-loan miswritings): a variant character borrowed purely for its sound (e.g. 訓 in 訓覺, standing in for 瞓) falls back to the variant character's own native reading unless the exact word is listed in data/variant_words.tsv. This is a curated, growing seed list (see CONTRIBUTING.md), not exhaustive — the set of such miswritings is open-ended and new ones surface over time.
• Fraction 十分之: Fractions with denominator 十 (e.g. 1/10) output fan1 instead of fan6 because 十分之 is also a common adverb (十分之好 = "extremely good") — the adverb entry in rime-cantonese takes longest-match priority. Other denominators (1/2, 1/3, 3/4, 1/12 …) produce the correct fan6 reading.
• Tone sandhi (變調): Citation tones only — tone sandhi is deferred to a future release.
• No neural polyphone tier: The current segmentation-based approach covers most cases. A BERT-based polyphone layer (trainable on HKCanCor CC-BY) is on the roadmap but not included in v1.

Comparison with other tools

canto-hk-g2p is the only tool that handles English code-switching cleanly — standard in Hong Kong Cantonese (e.g. 佢 send 咗 email 俾我) — and the only one with a Cantonese text normalizer for numbers, dates, and currency. It now also provides full IPA output with English tokens converted via the CMU Pronouncing Dictionary.

Data sources

Bundled in the wheel (all permissive)

Excluded (license-incompatible)

Build from source

Prerequisites

• Rust stable toolchain — install via rustup
• Python >= 3.8
• maturin >= 1.4

pip install maturin

Steps

# 1. Clone the repository
git clone https://github.com/typangaa/canto-hk-g2p.git
cd canto-hk-g2p

# 2. Fetch data sources (downloads ~20 MB: rime-cantonese, Unicode, CMU dict)
python3 scripts/fetch_data.py

# 3. Build binary pronunciation dictionaries
python3 scripts/build_dict.py

# 4. Build the Rust extension and install into the current environment
maturin develop --release
# Or build a wheel:
maturin build --release
pip install target/wheels/*.whl

Run tests

# Rust unit tests
cargo test

# Python integration tests (requires built extension + data/)
python3 -m pytest tests/ -v

All 190 Rust unit tests + 363 Python integration tests (553 total) should pass. The test suite covers basic G2P correctness, polyphone disambiguation, English passthrough, code-switching, punctuation normalisation, number/date/unit/currency/fraction/score normalization, batch processing, convert_detailed() / convert_detailed_batch() output structure, IPA conversion (all initials, finals, tones, syllabic consonants, CMU English lookup), user_dict runtime overrides, and the Candidates API (convert_candidates() / convert_candidates_batch()) including its confidence and source tags.

Contributing

Contributions are welcome. Please open an issue before starting significant work so we can align on direction. See CONTRIBUTING.md for coding guidelines, how to add entries to the hand-curated oral list, and the pull-request process.

A few areas where help is particularly valuable:

• Expanding data/oral_hk.tsv with additional colloquial characters
• Expanding data/variant_words.tsv with additional 借音字 (phonetic-loan) aliases
• Unit abbreviation expansion (km, °C, kg → Cantonese spoken form)
• Improving number normalization edge cases
• Adding tone sandhi rules (future release)
• Packaging and PyPI release CI

License

Apache-2.0 — see LICENSE.

Attribution (CC-BY-4.0 requirement)

This library bundles data derived from rime-cantonese (© CanCLID contributors), licensed under Creative Commons Attribution 4.0 International. See NOTICE for full attribution details.

Related projects

• SEA-G2P — upstream inspiration (Rust G2P for Vietnamese)
• ToJyutping — de-facto Python/JS Cantonese G2P front-end
• PyCantonese — comprehensive Cantonese NLP toolkit
• rime-cantonese — primary lexicon data source