TigreGotico/mirandese-ipa-dict-synthetic

Mirandese IPA Pronunciation Dictionary

A word-level IPA pronunciation dictionary for Mirandese (mwl), the Asturleonese language of Terra de Miranda, north-eastern Portugal, co-official in Portugal since 1999. To our knowledge this is the only machine-readable Mirandese IPA dictionary in existence, so it is built and documented to be trustworthy by construction: every row carries an explicit provenance tier, and the highest tiers are anchored to a native-speaker gold dictionary and to Leite de Vasconcelos's Estudos de Philologia Mirandesa (1900–1901).

This replaces the earlier mirandese_phonemizer_dataset.csv (a single-column synthetic draft). The lemma inventory and its metadata are retained; every transcription has been rebuilt and verified.

Columns

column	meaning
word	Mirandese headword, official orthography (Convenção 1999)
ipa	phonemic transcription in the single documented convention (see below)
ipa_sendim	Sendinês (mwl-x-sendim) form, only where it differs from ipa
ipa_ifanes	Ifanês / Raiano (mwl-x-ifanes) form, only where it differs
ipa_narrow	native-speaker narrow transcription, verbatim, where one exists
pos	part of speech (where known)
english, portuguese	glosses (where known)
dialect	primary variety tag: central / sendinês / raiano / all
provenance	trust tier — see below
vasconcelos_ref	page pin(s) in Estudos de Philologia Mirandesa I/II where the headform is attested
fable_corrections	per-row note when the transcription was corrected off the engine draft
notes, sources	free notes; which source lists contributed the lemma

The single documented convention (ipa)

The ipa column uses one convention for every row, the phonemic norm of the orthography2ipa Mirandese spec (documented in docs/languages/mwl.md):

• Sibilants — the medieval four-way contrast is kept: apico-alveolar s̺ z̺ (written with the retracted diacritic) vs. plain dorso-dental s z; ʃ ʒ for ⟨x⟩/⟨j,g⟩; the affricate t͡ʃ for ⟨ch⟩.
• Vowels — broad [u] (not the centralised [ʉ]); unstressed reduction a→ɐ, e→ɨ, o→u from the Portuguese adstrate; phonemic ɛ ɔ kept.
• Leonese features — word-initial ⟨l⟩ → ʎ (with the article/clitic bare-/l/ exception), ⟨lh⟩=ʎ, ⟨nh⟩=ɲ, diphthongs ⟨ie/iê⟩=jɛ, ⟨uo/uô⟩=wo; intervocalic /b/→β, /ɡ/→ɣ.

The narrow native-speaker realisations that fall outside this convention — centralised [ʉ], final [ʊ], spirant [ð], laminal [s̻ z̻], syllable dots and optional-phoneme parentheses — are not mixed into ipa; they are preserved verbatim in ipa_narrow.

Provenance tiers

Every row is tagged with exactly one tier, from most to least directly attested:

tier	n	meaning
native-dict	203	headword is in the native-speaker gold dictionary (TigreGotico/mirandese_g2p); its ipa is that expert human transcription rendered into the convention above, with the raw narrow form in ipa_narrow
attested-vasconcelos	569	headword (this exact spelling, accent-folded) is attested in Vasconcelos's Estudos de Philologia Mirandesa; ipa is the convention-checked engine transcription and vasconcelos_ref pins the page(s)
engine-convention	301	headword is not directly attested in either authority; ipa is the orthography2ipa engine transcription in the documented convention, honestly labelled as engine output
total	1,073

Three-source arbitration

The transcriptions were produced by a three-authority chain, mirroring the gold-20 trust methodology:

1. Engine draft — every lemma transcribed with orthography2ipa (dev 7.19.3a1) for mwl, mwl-x-sendim and mwl-x-ifanes. Benchmarked against the only human-authored Mirandese gold (TigreGotico/mirandese_g2p, 219 words), the engine scores ~19.5 % PER folded, ~13.1 % once the documented broad-vs-narrow notation gap is folded out — consistent with the 83/204 native-dict corrections below. (An earlier draft cited 0.00 % PER on the repo's 20-sentence-per-lect set; that set is engine-pinned — authored to match the engine — so 0 % was an internal consistency check, not human accuracy.)
2. Native-speaker gold (204 words) — where a headword exists in the expert-human mirandese_g2p dictionary, the human transcription is the authority: it supplies ipa_narrow verbatim and, rendered into the documented convention, overrides ipa. Of these, 120 agreed with the engine draft (phonemically, ignoring narrow-vs-broad notation) and 83 disagreed and were corrected to the native form (fable_corrections records the engine draft).
3. Vasconcelos & Convenção — Leite de Vasconcelos's Estudos de Philologia Mirandesa Vol. I (1900) and Vol. II (1901) provide the attestation anchor (657 headwords attested and page-pinned) and, together with the Convenção Ortográfica (1999) §L, the arbitration authority for the curated high-risk corrections (function-word/clitic initial-l de-palatalisation, etc.).

Verification statistics

• Native-dict words cross-checked against the engine: 204
• of which corrected to the native-speaker form: 83
• Headwords attested & page-pinned in Vasconcelos EPM I/II: 657
• Rows carrying a documented fable_corrections note: 88
• Rows with a distinct Sendinês variant: 282; with a distinct Ifanês/Raiano variant: 90

Sources of the lemma inventory

• The prior synthetic dataset's word list and its POS/gloss metadata.
• The native-speaker gold dictionary TigreGotico/mirandese_g2p (Central + Sendinês + Raiano entries).
• High-frequency vocabulary mined from the Mirandese Wikipedia (mwl.wikipedia.org) — a frequency list was built independently from a crawl of ~1,000 article extracts (no shared frequency corpus was available at build time); the highest-frequency types not already covered were added and transcribed as engine-convention.

Honest limitations

• attested-vasconcelos means the headword spelling is attested in EPM; it is not a claim that Vasconcelos gives that exact IPA (his 1900 notation is not IPA). Vasconcelos writes in a pre-Convenção orthography, so many modern spellings legitimately do not match and fall to engine-convention.
• engine-convention rows are engine output, convention-checked but not individually human-verified.
• The Sendinês/Raiano columns are largely engine-derived; the native-speaker gold for those sub-dialects is small (n≈11 Sendinês).
• Mid-vowel height (e/ɛ, o/ɔ) and stress placement on engine-convention rows carry the engine's residual error.

Citation

• Ferreira, M. B. & Raposo, D. (coords.) (1999). Convenção Ortográfica da Língua Mirandesa. Câmara Municipal de Miranda do Douro / CLUL.
• Leite de Vasconcelos, J. (1900, 1901). Estudos de Philologia Mirandesa (Vols. I–II). Imprensa Nacional, Lisboa.
• Native-speaker gold: TigreGotico/mirandese_g2p.
• Engine: orthography2ipa (TigreGotico).