MODEL 06 · IPA2VEC · PHONETIC EMBEDDINGS

Soundalike Search

      
      Powered by MondoPhone (IPA2vec embeddings on phonetic output)
    

Find names that sound like yours across world languages. The engine behind echoes.mondonomo.ai uses IPA2vec — articulatory-feature embeddings on PolyIPA's phonetic output — to cluster names across scripts by phonetic similarity. The same primitive powers KYC entity-resolution: Yousef, Yusuf and يوسف are the same person.

Model card Try echoes.mondonomo.ai ↗

Try it — name → soundalikes across languages

Enter a name. The model returns the closest phonetic matches in 6+ languages, ranked by IPA2vec cosine similarity.

EXAMPLES:

Enter a name and press Find echoes.

Model card

Architecture: IPA2vec — fastText on character n-grams of IPA strings, articulatory-feature regularized
Helper model: similarIPA — handles phonetic notation variations (broad vs narrow transcription)
Embedding dim: 128
Distance: Cosine, calibrated to panphon articulatory-feature distance
Corpus: 1M Nomograph names × 20 sampled languages → ~20M (name, IPA) pairs
Latency: ~12ms for top-50 (ANN-indexed)
Use cases: KYC matching · echoes search · fuzzy dedup · name discovery

API

curl -X POST https://api.mondonomo.ai/v1/soundalikes \
  -H "Authorization: Bearer $TOKEN" \
  -d '{
    "name": "Yusuf",
    "languages": ["en","fr","de","ru","ja","th","zh","hi"],
    "k": 10
  }'

{
  "query": "Yusuf",  "query_ipa": "/ˈjuː.suf/",
  "matches": [
    {"name": "Joseph",  "lang": "en",
     "ipa": "/ˈdʒoʊzəf/",  "sim": 0.73},
    {"name": "Youssef", "lang": "fr",
     "ipa": "/jusɛf/",   "sim": 0.91},
    ...
  ]
}

How it relates to PolyIPA. Soundalike sits one layer above PolyIPA in the stack: Name → PolyIPA → IPA → IPA2vec → 128-dim vector → ANN search → ranked candidates. Because the embedding is trained on articulatory features (not raw characters), it generalizes across writing systems that share no surface form — Joseph ↔ يوسف ↔ Йосиф all cluster together.

MODEL 01

PolyIPA

The G2P engine that produces the IPA strings IPA2vec embeds.

g2p · p2g→

MODEL 02

Transliterator

When a corpus match exists, prefer corpus-weighted transliteration over phonetic similarity.

scripts→

MODEL 04

Name Parser

Soundalike-match on each parsed slot independently for surgical KYC matching.

structured→

Soundalike Search

Try it — name → soundalikes across languages

Model card

API

Built on the same pipeline.

PolyIPA

Transliterator

Name Parser