One name in. Any script. Out.

MondoPhon is the universal proper-name-to-proper-name mapping engine — an ensemble of Transformer and Weighted Finite-State Transducers covering every script in active use. Romanization, deromanization, phonetic transcription, G2P, P2G — all under one API.

Try PolyIPA (P2G live) → PolyIPA paper ↗

TASK SURFACE

Five mappings, one ensemble.

Every script-to-script and script-to-sound conversion you need to handle the world's names. Where a learned Transformer wins (named entities, soft phonetic rules), MondoPhon uses it. Where a WFST wins (deterministic Romanization standards, compose-able cascades), it uses that. The ensemble routes per-task per-language.

TASK 01 · ROMANIZATION

Romanize

尤金 → Yóujīn
ณัฐกรณ์ → Natthakon
يوسف → Yusuf

RUN DEMO → TASK 02 · DEROMANIZATION

Deromanize

Yusuf · ar → يوسف
Eugene · ru → Юджин
Kenta · ja → 健太

RUN DEMO → TASK 03 · G2P

Grapheme → IPA

Eugen · de → /ˈɔʏɡn̩/
ณัฐกรณ์ · th → /nát̚.tʰa.kɔːn/

RUN DEMO → TASK 04 · P2G

IPA → Grapheme

/ˈjuːdʒɪn/ · en → Eugene · Eugen · Eugéne
Top-3 with beam-90 search

RUN DEMO → TASK 05 · SOUNDALIKES

Phonetic similarity

Yusuf ≈ Joseph ≈ Youssef ≈ Yosef
IPA2vec cosine over 128-dim embeddings

RUN DEMO → TASK 06 · COMPOSE

Cross-script chains

ar → IPA → Latn → IPA → ja
Multi-hop conversion with confidence at every step

SEE BUSINESS CASE →

LEGACY MODELS · THE SCIENCE BEHIND

Two production papers feed MondoPhon today.

The full ensemble is in preview, but its production components have been published, evaluated and are live on the API. They demonstrate the research lineage that MondoPhon will unify into one endpoint.

DEC 2024 · arXiv:2412.09102

PolyIPA — Multilingual Phoneme-to-Grapheme Conversion

Davor Lauc · the P2G shipping today

CER 0.055 char-BLEU 0.914 top-3 CER 0.026 ByT5-small · 300M params

The current production P2G inside MondoPhon. ByT5-based seq2seq, trained on WikiPron + 1M Nomograph augmentations. Companion helper models IPA2vec and similarIPA also power the soundalike search.

Try the demo →

DEC 2024 · arXiv:2412.03877

AyutthayaAlpha — Thai-Latin Transliteration Transformer

Lauc · Rutherford (Chulalongkorn) · Wongwarawipatr

CER 0.0047 82.32% top-1 95.24% top-3 2.7M Thai-Latin pairs

The Thai romanization path in MondoPhon today. ByT5-based, with a surprisingly competitive AyutthayaAlpha-VerySmall variant. Published with Chulalongkorn — the corpus comes from the Handbook of Top Thai Names.

Try Thai romanization →

ENSEMBLE ARCHITECTURE

Transformer where it wins. WFST where it wins.

Naïve character-level transliteration breaks on every irregular form (silent letters, tones, kunya). Pure neural models hallucinate when scaled to 150+ scripts. MondoPhon routes per-task per-language: deterministic Romanization standards (RTGS, ALA-LC, BGN/PCGN, Hepburn, Pinyin) run as WFST cascades for speed and auditability; named-entity-aware mappings and phonetic prediction run on the Transformer ensemble. At inference, the router decides which path to take based on input language, target script, and confidence.

Try the transliterator → See soundalikes →

# Compose a chain: ar → IPA → ja
curl -X POST https://api.mondonomo.ai/v1/mondophon \
  -H "Authorization: Bearer $TOKEN" \
  -d '{
    "name": "يوسف",
    "source_script": "arab",
    "chain": ["g2p", "p2g"],
    "target_script": "jpan"
  }'

# →
{
  "output": "ユースフ",
  "steps": [
    {"task":"g2p", "out":"/ˈjuː.suf/", "conf":0.94},
    {"task":"p2g", "out":"ユースフ", "conf":0.87}
  ],
  "engine": "polyipa+wfst"
}