say-my-name

Personal names paired with their IPA pronunciation in 8 languages. Built to train and benchmark grapheme-to-phoneme (G2P) models on names specifically, the case general G2P and TTS systems handle worst. Part of the say-my-name project.

lang is the pronunciation language, not the name's origin. Each row comes from a Wiktionary language section, so it records how this name is said in that language — "Michael" appears as de ˈmɪçaˌeːl, fr mikaɛl and it ˈmajkol. Michael's etymological origin is Hebrew, and no row here encodes that. About 600 spellings appear under more than one language for exactly this reason.

Format

TSV, UTF-8, one row per attested pronunciation (a name with several documented pronunciations appears on several rows):

column	meaning
name	the name, in its native script
lang	ISO 639-1 pronunciation language (ar, de, es, fr, it, pl, tr, vi)
kind	given, surname, or both
ipa	normalized IPA (NFC; stress and length marks kept)

Files: a combined names_ipa.tsv and splits/ with train.tsv / dev.tsv / test.tsv. (The per-language intermediates ar.tsv … vi.tsv are build artifacts, not part of the release; regenerate them from the repo with python -m pronounce.extract.)

Splits

Split by (lang, name) pair, disjoint, seed=13: a name in a given language appears in exactly one split, with all of its pronunciation variants, so evaluation measures generalization to unseen names rather than dictionary recall.

The key is the pair, not the bare spelling. ~116 spellings (Maria, Anna, …) exist as names in more than one language here, so they can appear in train under one language and in test under another — different language tag, different target. "Disjoint by (lang, name)" is the accurate phrasing; an earlier version of this card said "by name", which claims slightly more.

split	names	rows (name, IPA)
train	20,084	24,841
dev	2,507	3,120
test	2,507	3,091

Per-language sizes

lang	names	variants
ar	575	617
de	1,705	2,085
es	2,909	7,199
fr	2,570	2,851
it	2,581	2,739
pl	14,071	14,241
tr	317	359
vi	370	961
total	25,098	31,052

Source & build

Mined from Wiktionary via Wiktextract dumps on kaikki.org: fetch → filter to name entries with an IPA transcription → normalize IPA → split. Reproducible from the repo (make data).

Limitations

• Heavily imbalanced: Polish is ~56% of the data; Turkish, Vietnamese and Arabic are small (300–600 names). Models trained on this inherit that skew, and any micro-averaged score over the whole set is largely a Polish score — report a macro-average alongside it.
• Transcription style varies by language, and that is not a pronunciation difference. French and Vietnamese entries carry no stress marks at all, while Italian, Spanish, Polish and Turkish ones essentially always do; Polish writes tie-barred affricates (t͡ʂ) where many G2P systems emit plain digraphs. Comparing a system's output against these references without reconciling that first produces 0% scores that look like pronunciation failures and aren't — the repo's pronounce.stress and metrics.align_stress_convention exist for exactly this.
• Polish IPA on Wiktionary is largely template-generated from spelling (1.01 variants per name, against 2.47 for Spanish). Treat very high Polish scores as evidence about a transliteration rule, not about names.
• Pronunciations are whatever Wiktionary attests; coverage varies by language and no dialect unification is applied.
• 81 German rows that were syllable fragments rather than pronunciations (ˈbaɐ̯-, -bə-, -ˌʁaː — Wiktionary transcribing "Barbara" piece by piece) were removed. No name lost all of its variants, so the splits are unchanged.

License

Wiktionary content is CC BY-SA; this derived dataset keeps CC BY-SA 4.0. Project code is MIT.

Links

• Model: saraelhark/say-my-name-byt5-small
• Code: github.com/saraelhark/say-my-name