thaiphon 0.5.0

Thai phonology engine with pluggable transliteration systems.

thaiphon

A zero-dependency Thai phonological transliteration engine. 75 % exact-match accuracy against independent Wiktionary IPA ground truth when installed with the optional thaiphon-data-volubilis lexicon package; 57 % on the base engine alone (see Accuracy for details).

• Try it online: rianthai.pro/thai-transliteration — no install needed. Built for teachers, Thai language learners, and anyone who just wants to type a word and see its pronunciation.
• Documentation: rianthai.pro/docs/thaiphon

What it does

Thai script goes in; transliteration comes out. The engine first resolves the input into a phonological word — a tuple of syllables, each decomposed into IPA-based onset, vowel quality, vowel length, coda, and tone. Every output scheme is then a small declarative SchemeMapping that renames those phonemes into a target surface form. Add a new scheme by writing data, not code.

Eight built-in renderers ship with the package:

• ipa — IPA with tone contour digits (the default phonological representation used for the Wiktionary benchmark).
• tlc — thai-language.com "Enhanced Phonemic" with bracketed tone tags ({M}, {L}, {H}, {F}, {R}). Supports format="html", which emits the tone markers as <sup> tags instead of braces (naam<sup>H</sup>) to match the thai-language.com rendering.
• rtgs — Royal Thai General System of Transcription (the 2002 Royal Institute specification). Plain ASCII Latin output with no tone marks, no length distinction, and no diacritics — aspirated stops as ph, th, kh, ch; six final consonants (k, t, p, n, m, ng). The canonical "official" form used on road signs, map names, and government publications.
• rtl — Rak Thai Language School convention. IPA-style vowels (ʉ ɛ ɔ ə), aspirated-stop digraphs (ph, th, kh, ch), an explicit ʼ (U+02BC) for vowel-initial syllables, and a macron on mid-tone vowels. Syllable separator is a single space.
• paiboon — the original Paiboon Publishing romanization used in the first-edition Thai for Beginners series. Aspirated stops take bare letters (p, t, k, ch); unaspirated voiceless stops use English-cluster digraphs (bp, dt, g, j). IPA-style vowels double for length (aa, ii, ʉʉ). Centring diphthongs have no short/long distinction in spelling (ia, ʉa, ua). Syllable separator is -.
• paiboon_plus — the revised system adopted in the 2009 Three-Way dictionary. Identical to paiboon except the centring diphthongs double for length (short ia ʉa ua, long iia ʉʉa uua) and the triphthong-like combinations follow suit (iiao, uuai, ʉʉai).
• morev — Cyrillic transliteration following the Russian-academic dictionary-citation tradition. Aspirated stops as digraphs (кх, тх, пх) in text mode and as к<sup>х</sup> etc in HTML mode. Vowel length marked by a combining macron; tone by a spacing diacritic at the end of the syllable.
• lmt — learner-textbook Cyrillic in the Lipilina-Muzychenko-Thapanosoth convention. Tones render as Unicode superscript digits (⁰ mid, ¹ low, ² falling, ³ high, ⁴ rising); vowel length is marked with an ASCII colon (ка: for /kaː/); syllables inside a word are space-separated. HTML mode wraps the tone digit in <sup>…</sup>. Complements morev with a textbook-style layout rather than dictionary-citation typography.

Custom schemes are straightforward to register via SchemeMapping — see Custom schemes.

Install

uv add thaiphon thaiphon-data-volubilis
# or
pip install thaiphon thaiphon-data-volubilis

Requires Python 3.10+. The engine itself has no runtime dependencies. thaiphon-data-volubilis is a companion package shipping a ~35 k-entry Thai lexicon derived from the VOLUBILIS Mundo Dictionary. thaiphon detects it automatically on import and uses it for word-boundary segmentation and word-specific readings that a rule-based engine can't infer from orthography alone.

The data package is CC-BY-SA 4.0, separate from the Apache-2.0 engine — installing it doesn't change the license of your own code.

Minimal install (not recommended for most users)

If there's a hard reason to keep dependencies lean:

pip install thaiphon

The engine still runs, but you give up a lot:

• Accuracy on the Wiktionary IPA etalon drops from ~75 % to ~57 %.
• Common words break in visible ways. ส้ม ("orange") comes out as /sa˥˩.ma˦˥/ — two syllables — when the correct reading is /som˥˩/, one closed syllable. The rules alone can't choose between "closed syllable with inherent /o/" and "two open syllables with inserted /a/"; the lexicon settles that kind of ambiguity.
• Sanskrit-learned compounds miss their learned readings. มหาวิทยาลัย loses the /tʰa/ insertion between /wit̚/ and /jaː/. The spoken form is /wit̚.tʰa.jaː.laj/; the base engine emits /wit̚.jaː.laj/.
• Failures are silent. The engine always returns a transliteration — no error, no flag, no way to tell at the output layer which of the ~43 % of dictionary entries fell through. A caller can't filter out bad results after the fact.

For anything user-facing, indexing, or searching, install both packages.

Production footprint

thaiphon runs comfortably in long-lived web processes, container orchestrators, and serverless runtimes. The engine is pure Python with no heavyweight imports. The optional lexicon (when installed) ships as a read-only SQLite database, memory-mapped at first use.

Measurement	Engine alone	+ thaiphon-data-volubilis
RSS after first transcribe_sentence	~10 MiB	~19 MiB
RSS after 1 k lookups	~10 MiB	~47 MiB
Median lookup latency (lexicon hit)	—	~13 µs
Cold-start import cost	negligible	negligible

The ~30 MiB of additional resident set after warmup is a bounded per-process LRU over the most recently inflated entries. The bulk of the lexicon stays in mmap'd pages of lexicon.db, counted once by the operating system regardless of how many Python workers have imported the package.

Web servers (Django, FastAPI, Flask)

Run as many workers as your CPU budget allows. The lexicon does not multiply with worker count, because the OS page cache shares the mapped database across processes:

• gunicorn -w 8 myapp.wsgi → roughly 8 × 50 MiB of Python objects plus one shared ~25 MiB for the lexicon. Not 8 × 350 MiB.
• uvicorn --workers 4 myapp:app behaves the same for FastAPI and Starlette. Inside one worker, async handlers share a SQLite connection-per-thread; the threadpool gets one connection per active thread, opened on first use.
• Django works in sync views, async views, and runserver without any setting changes.

The lexicon loads lazily. The first request that touches a Thai word pays the SQLite open and first index-page read (single-digit milliseconds). Subsequent requests within that worker hit the LRU.

Serverless (AWS Lambda, Cloud Run, Fargate)

Cold starts are dominated by your Python interpreter and your own imports. thaiphon adds almost nothing at import time, since the lexicon is only read from on first lookup. No eager inflation pass, no .pyc of several hundred MiB to unpack. The 512 MiB Lambda tier is plenty even with thaiphon-data-volubilis installed; the lexicon's footprint is the on-demand LRU, not the full ~25 MiB database.

Both packages fit comfortably in a Lambda layer.

Multi-process scripts (multiprocessing, joblib, Dask)

Worker processes spawned via fork() (Linux/macOS default) and spawn() (Windows, or after set_start_method) both work. The SQLite connection is held in a threading.local() slot, so it never crosses the fork boundary. Each child opens its own connection on first use; all of them read from the same mmap'd file.

Tests under pytest-xdist

Each xdist worker is a separate Python process. With the lexicon installed, per-worker RSS after warmup is in the tens of MiB. Running -n 8 on a laptop is uneventful.

Quick start

from thaiphon import (
    transcribe, transcribe_word, transcribe_sentence, analyze, list_schemes,
)

# Discover available schemes.
list_schemes()
# ('ipa', 'lmt', 'morev', 'paiboon', 'paiboon_plus', 'rtgs', 'rtl', 'tlc')

# IPA — the phonological representation.
transcribe("ภูมิ", scheme="ipa")
# '/pʰuːm˧/'

# TLC — thai-language.com convention (braces by default; <sup> in HTML).
transcribe("ลิฟต์", scheme="tlc")
# 'lif{H}'
transcribe("ลิฟต์", scheme="tlc", format="html")
# 'lif<sup>H</sup>'

# RTGS — the 2002 Royal Institute "official" romanization.
transcribe("กรุงเทพ", scheme="rtgs")
# 'krungthep'
transcribe("ลิฟต์", scheme="rtgs")
# 'lip'  (strict native-coda collapse; no tone, no diacritics)

# Paiboon / Paiboon+ — textbook romanizations.
transcribe("เสื้อ", scheme="paiboon")
# 'sʉ̂a'
transcribe("เสื้อ", scheme="paiboon_plus")
# 'sʉ̂ʉa'

# RTL School convention.
transcribe("ภาษาไทย", scheme="rtl")
# 'phāa sǎa thāy'

# Morev — Russian-academic dictionary-citation Cyrillic.
transcribe("สวัสดี", scheme="morev")
# 'саˆ-ватˆ-дӣ'

# LMT — learner-textbook Cyrillic with superscript-digit tones.
transcribe("สวัสดี", scheme="lmt")
# 'са¹ ват¹ ди:⁰'

# Single-word shortcut that skips sentence tokenization.
transcribe_word("สวัสดี", scheme="ipa")
# '/sa˨˩.wat̚˨˩.diː˧/'

# Sentence segmentation + rendering.
transcribe_sentence("ฉันชอบกินข้าว", scheme="ipa")
# '/t͡ɕʰan˩˩˦/ /t͡ɕʰɔːp̚˥˩/ /kin˧/ /kʰaːw˥˩/'

# Reading profile switches colloquial vs citation pronunciation — see below.
transcribe("ลิฟต์", scheme="ipa", profile="etalon_compat")
# '/lip̚˦˥/'

# Access the intermediate phonological word.
result = analyze("ผลไม้")
for syl in result.best.syllables:
    print(syl.onset, syl.vowel.symbol, syl.vowel_length.name, syl.coda, syl.tone.name)

Reading profiles

The profile argument controls how the engine handles foreign phonotactics in loanwords, especially the coda consonants that native Thai normally neutralises:

# "Lift" / elevator — final /f/ from English.
transcribe("ลิฟต์", scheme="ipa", profile="everyday")         # '/lif˦˥/'
transcribe("ลิฟต์", scheme="ipa", profile="etalon_compat")    # '/lip̚˦˥/'

# "Graph" — register-sensitive.
transcribe("กราฟ", scheme="ipa", profile="everyday")          # '/kraːp̚˨˩/'
transcribe("กราฟ", scheme="ipa", profile="careful_educated")  # '/kraːf˨˩/'

The four supported profiles:

• everyday (default) — colloquial urban pronunciation. Preserves foreign codas in well-integrated modern loans (ลิฟต์, เคส, อีเมล) and collapses them in older or less register-sensitive words (กราฟ, กอล์ฟ, บัส).
• careful_educated — broadcast/formal register. Preserves more foreign codas than everyday.
• learned_full — restores full Sanskrit/Pali learned readings for Indic-derived words (ภูมิ, ปกติ).
• etalon_compat — dictionary-citation style; collapses every foreign coda to its native-Thai equivalent. Useful for matching pronunciation dictionaries.

Preservation is per-word, driven by the lexicon and attested usage — not by blanket rules that paper over lexical convention.

Accuracy

Measured by exact-match rate against independent, publicly-licensed Thai phonology corpora:

Corpus	License	Size	Scheme	Base engine	With thaiphon-data-volubilis
kaikki.org Thai Wiktionary	CC-BY-SA 4.0	17,014	ipa	~57 %	~75 %
PyThaiNLP G2P (Wiktionary)	CC0	15,782	ipa	—	~73 %

"Base engine" is pip install thaiphon on its own. The jump to 75 % comes from the thaiphon-data-volubilis lexicon (see Install) — the word-boundary and variant coverage a rule-based engine can't infer from orthography alone.

Both numbers come from public data. The primary kaikki.org etalon and PyThaiNLP's independent G2P extraction are cross-checks on each other; if the engine breaks in a way that only one source catches, the other will usually catch it too.

Reproducing the numbers yourself:

The repository ships a 2,500-entry random sample of the kaikki.org Thai Wiktionary dump (seed 20260421, CC-BY-SA 4.0) as a bundled pytest fixture. Install the data package, then run the etalon tests:

# Ensure both packages are installed (see Install section).
pip install thaiphon thaiphon-data-volubilis

# Bundled sample — runs in seconds, no external download:
pytest tests/etalon/test_wiktionary_ipa_sample.py -v
# Floor: 72 %.  Measured: ~74 % on the sample.

# Full 17,014-entry measurement — download the kaikki.org dump first:
#   https://kaikki.org/dictionary/rawdata.html  (Thai entries, ~43 MB)
#   Place at ~/.cache/thaiphon/kaikki-thai.jsonl  (or set $THAIPHON_KAIKKI)
pytest tests/etalon/test_wiktionary_ipa_full.py -v
# Floor: 73 %.  Measured: ~75 % on the full corpus.

If thaiphon-data-volubilis is not installed, both etalon tests skip with a message pointing at the install command, so make test always finishes cleanly whatever your setup. See tests/README.md for more detail and tests/fixtures/README.md for fixture licensing and sampling parameters.

Custom schemes

from thaiphon.model.enums import VowelLength
from thaiphon.renderers.mapping import SchemeMapping, MappingRenderer
from thaiphon.registry import RENDERERS
from thaiphon import transcribe

pedagogical = SchemeMapping(
    scheme_id="ped",
    onset_map={"kʰ": "kh", "k": "k", "tɕ": "j", "m": "m", "n": "n", ...},
    vowel_map={
        ("a", VowelLength.SHORT): "a",
        ("a", VowelLength.LONG):  "aa",
        ...
    },
    coda_map={"m": "m", "n": "n", "ŋ": "ng", "p̚": "p", "t̚": "t", "k̚": "k"},
    tone_format=lambda base, syl: base,   # no tone decoration
    syllable_separator="-",
)
RENDERERS.register("ped", lambda: MappingRenderer(pedagogical))

transcribe("สวัสดี", scheme="ped")

Architecture

Key design decisions:

• Zero runtime dependencies. The full phonology, rules, and lexicon ship as plain Python with no C or FFI.
• Immutable phonological model via frozen dataclasses.
• Lexicons (loanword overrides, Indic learned readings, royal vocabulary, etc.) are module-level Python literals — auditable and grep-able.
• Scheme-specific editorial conventions live in the renderer layer, not the phonology. Improvements to onset/vowel/coda derivation show up in every output scheme at once.

Development

git clone https://github.com/5w0rdf15h/thaiphon
cd thaiphon
make dev          # install with dev dependencies via uv
make test         # pytest -q
make lint         # ruff
make typecheck    # mypy src/thaiphon
make check        # test + lint + typecheck
make format       # black + isort

The test suite is organised in two layers:

• tests/*.py — public API surface and representative regression examples. Fast, deterministic, and doubles as living API documentation.
• tests/rules/*.py — internal phonological-rule coverage (tone matrix, coda collapse, onset resolution, cluster strategies, lexicon membership, etc.). Exercises internal modules directly; the right place to add a regression when fixing a derivation bug.

All example words in tests are hand-curated from openly-licensed sources only — Wiktionary (CC-BY-SA 4.0) and VOLUBILIS Mundo Dictionary (CC-BY-SA 4.0). See tests/fixtures/README.md for the full provenance policy.

Credits

Includes code adapted from the PyThaiNLP project (Apache-2.0). See NOTICE for the full attribution. Optional lexicon data in thaiphon-data-volubilis derives from the VOLUBILIS Mundo Dictionary (CC-BY-SA 4.0).

Acknowledgements

Personal thanks to the teachers and principal at RTL School — Rak Thai Language School for teaching and explaining Thai to me. This engine's phonological intuitions come directly from their patient, careful instruction.

License

Apache-2.0. See LICENSE.