Languages and pronunciation

This page sets out, for each language in the library, what the sound row on the page claims, what the code says it rests on, and where it is weakest. It is written for specialists deciding whether to trust the pronunciation row for their language and how to correct it. Every statement here is taken from the module in pipeline/weft/, its data files, or the manifests and notes of the works that use it. Where the repository names no source, this page says so.

What the sound row is

How to propose a change

A pronunciation change is made only with a citation: a grammar, an article or a corpus.

  1. The Pronunciation issue form (.github/ISSUE_TEMPLATE/pronunciation.yml), linked from each page with the work and scheme filled in. It asks for the scope (one word, a rule, the stress or accent rule, or a new scheme alongside the existing ones), the words affected, what the sound should be (IPA if possible), the citation, and how you wish to be credited. No git is needed.
  2. A pull request against pipeline/weft/<module>.py and its data files under pipeline/weft/data/. For languages where the reading is entered per word (Egyptian, Akkadian, Sumerian, Vedic, Tamil, Persian, the runes, Japanese), the per-word form lives in the work's edition file, texts/<work>/edition.yaml. After a rule change, run weft acquire and weft draft for each work in that language, then weft check <work> and uv run --with pytest pytest -q tests before submitting. weft say <Greek words> prints Greek words in the restored and Erasmian schemes; other languages have no command-line phonemizer yet.

A different reconstruction can be added as a new scheme rather than replacing the current one; schemes sit side by side and the reader chooses.

Summary

"First" marks the page default. Confidence is the manifest's sound_confidence; at none the page shows no invitation.

LanguageModuleWorksSchemesConfidence
Old EgyptianegyptianPyramid Texts of Unasold-egyptian (first), egyptologicallow
SumeriansumerianEnheduanna, Temple Hymn 5recon-2300 (first), classroomlow
AkkadianakkadianHammurabi, prologueob-1750 (first), classroomlow
Vedic SanskritsanskritRigveda 1.1vedic (first), modernmedium
Ancient GreekgreekIliad, Odyssey, Plato, Aristotlerestored (first), erasmianmedium
Nicene Creed (381)koine (first), erasmianmedium
John, Beatitudes, 1 Corinthians 13koine (first), erasmianmedium
Epictetus, Marcus Aureliuskoine (first), restored, erasmianmedium
Biblical HebrewhebrewGenesistiberian (first), modern-israelimedium
The Shema (Deuteronomy 6:4-9)tiberian (first), modern-israelimedium
Biblical AramaichebrewDaniel 5tiberian (first), modern-israelimedium
Jewish Aramaic (liturgy)hebrewThe Kaddishtiberian (first), modern-israelilow
Classical ChinesechineseSunzi, Daodejingold-chinese (first), tang, mandarinmedium
Li Baitang (first), mandarin, cantonesemedium
Sahidic CopticcopticGospel of Mark 1sahidic (first), bohairicmedium
LatinlatinOvid, Res Gestaeclassical (first), ecclesiasticalmedium
Bede, Ecclesiastical History I.15 (731)classical (first, as Northumbrian book Latin), ecclesiasticalmedium
Bayeux Tapestryanglo-norman (first), classicalmedium
Henry I's Charter of Libertiesanglo-latin (first), classicalmedium
Magna Cartaanglo-latin (first), classicalmedium
Saxo, Gesta Danorum 9.3-5danish-medieval (first), classical, ecclesiasticalmedium
Saer de Quincy, two chartersanglo-latin (first), classicalmedium
Picoitalian-humanist (first), classicalmedium
Inter caeteraitalian-humanist (first), classicalmedium
Erasmuslow-countries (first), classicalmedium
More, Utopiatudor-english (first), classicalmedium
Luther, Ninety-five Thesesgerman-humanist (first), ecclesiastical, classicalmedium
Descartes and Newtonas-first-read (first), classicalmedium
Old TamiltamilTirukkuralold-tamil (first), modernmedium
RunicrunicKylver, Gallehus, Rökas-carved (only)low
Old IrisholdirishPangur Bánold-irish (only)low
Middle WelshwelshY Gododdin, Pa gur, Preiddeu Annwnmiddle-welsh (only)low
Old EnglisholdenglishBeowulf, Cædmon's Hymnwest-saxon (only)medium
Old NorsenorseVöluspá, Hávamál, Vafþrúðnismál, Grímnismál, Lokasenna, Þrymskviða, Grettis sagaold-norse (first), modern-icelandicmedium
Snorri, Gylfaginningold-norse (first), modern-icelandicmedium
Old East SlavicoldeastslavicPrimary Chronicle, 859-862orv1100 (first), rulow
PersianpersianRubaiyatearly (first), modernmedium
Middle MongolianmongolianSecret History of the Mongols, 1-10mm-1250 (only)low
JapanesejapaneseBashōedo-1686 (first), modernmedium
Franco-ItalianoldfrenchMarco Polo, Cipangufr1300 (first), it1300medium
ItalianitalianDante, Petrarch, Machiavelliflorentine (first), modernmedium
Da Ponte, Le nozze di Figaro (1786)florentine (first), modernmedium
Verdi, Brindisi and Credo (1853, 1887)florentine (first), modernmedium
SpanishspanishColumbus, 1493c1492 (first), modernmedium
PortugueseportugueseCamões, Os Lusíadas I.1-3lisboa1540 (first), europeu, brasileiromedium
Caminha, letter of 1500lisboa1540 (first), europeu, brasileiromedium
DutchdutchLinschotenh1596 (first), modernmedium
FrenchfrenchMontaignem1580 (first), modernmedium
Declaration of the Rights of Man (1789)fr1791 (first), modernmedium
de Gouges, Rights of Woman (1791)fr1791 (first), modernmedium
Rouget de Lisle, La Marseillaise (1792)fr1791 (first), modernmedium
Napoleon, farewell to the Old Guard (1814)fr1791 (first), modernmedium
Berlin Act, 1885fr1885 (first), modernmedium
GermangermanKant, Nietzschenorthern (first), modernmedium
Marx and Engels, Manifest (1848)northern (first), modernmedium
Einstein, inertia and energy content (1905)northern (first), modernmedium
Schikaneder, Die Zauberflöte (1791)northern (first), modernmedium
Luther's Bibleecg1545 (first), modernmedium
SwahiliswahiliSteere, The Kites and the Crowsz1870 (first), modernmedium
CherokeecherokeeConstitution of the Cherokee Nation (1827), the preamble and Article Ic1828 (first), modernlow
EsperantoesperantoZamenhof, Unua Libro specimens (1887)zamenhof (only)none
Quenya, Sindarinelvishprivate works only (Tolkien, in copyright)tolkien (only)none

The registry that maps a manifest's language code to a module is PHON in pipeline/weft/draft.py.

Old Egyptian

Schemes.

What it rests on. The consonant values follow "the outline of Loprieno (1995) and Allen (2013)" (module docstring; the notes cite Loprieno, Ancient Egyptian: A Linguistic Introduction, chapter 3, and Allen, The Ancient Egyptian Language: An Historical Study). The ejective reading of d and ḏ is attributed to Loprieno. Each of the fourteen vocalized forms names its evidence in the vocalization map of texts/pyramid-texts-unas/edition.yaml (for example Coptic ⲛⲟⲩⲧⲉ for nā́ṯir, Greek Ὄννος for the king's name), and the page shows it in the word's provenance. The module states that these are Weft's inferences by the sound correspondences in those grammars, not forms quoted from them.

Where it is weakest. The module calls this "the most reconstructed sound row in the library" and "conjectural word by word". Only fourteen forms are vocalized; every other word falls back to the Egyptological convention inside the first scheme, in lower case without stress, and is flagged per token. Each consonant value in the list above is marked as debated; the key notes that ꜣ may have weakened to a glottal stop or y by the Middle Kingdom and that some scholars take ꜥ as a d-like stop in the earliest texts. Several vocalization entries call individual vowels uncertain or a guess.

What a specialist could improve. Add or correct entries in the vocalization map and the n field of the tokens, with the evidence named; revise consonant values in egyptian.py; propose an alternative consonant system as a second reconstruction.

Sumerian

Schemes.

What it rests on. Jagersma, A Descriptive Grammar of Sumerian (2010), "in broad agreement with Edzard, Sumerian Grammar (2003)" (module docstring). The notes add Foxvog, Introduction to Sumerian Grammar, on pronunciation. Lemma and morphology come from the Oracc lemmatization (epsd2/literary); the cuneiform row from the Oracc Sign List value table.

Where it is weakest. The module describes the language as an isolate whose sound "is known only indirectly" and the row as "among the most reconstructed" in the library. Stress and vowel length are unknown and not shown. The phoneme conventionally written dr or ř is not applied to any word (manifest and module). One broken sign has no sound.

What a specialist could improve. The rules in sumerian.py (stop series, the treatment of written doubling, whether and where to apply ř); per-token readings in the optional n field of texts/enheduanna-temple-hymns/edition.yaml where the spelling misleads (the module's example is a-a for aya).

Akkadian

Schemes.

What it rests on. The sibilant view is attributed to "Streck and Kogan among others" (module); the notes cite M. P. Streck, "Sibilants in the Old Babylonian texts of Hammurapi and of the governors in Qaṭṭunān" (2006). Stress in both schemes follows the rule in Huehnergard, A Grammar of Akkadian (3rd edition, 2011). The sound is computed from a hand normalization (n) in the edition; the cuneiform row is generated from a modern sign reading (sg) through the Oracc Sign List (pipeline/weft/data/akk_osl_values.tsv, CC0).

Where it is weakest. The manifest: the values of š, s, z and the emphatics are debated, and stress is inferred by a modern rule. The key says the exact sound of the emphatics is not known. The transliteration is Harper's of 1904, with older conventions; the sound does not depend on it, since it reads n.

What a specialist could improve. The normalizations in texts/akkadian-hammurabi/edition.yaml (vowel length and contraction drive both sound and stress); the consonant values and the stress rule in akkadian.py; a second reconstruction that keeps š as sh, which the notes name as the competing view.

Vedic Sanskrit

Schemes.

What it rests on. The ancient phonetic treatises (prātiśākhyas) and Pāṇini's statement on short a (module). The notes cite W. S. Allen, Phonetics in Ancient India (1953), and Macdonell, A Vedic Grammar for Students (1916), appendix III; and Arnold, Vedic Meter (1905), for the meter. The accent is hand-entered in the IAST form n; decode_marks reads the Devanagari accent strokes back into raised syllables so the two can be checked against each other (tested in tests/test_golden.py).

Where it is weakest. The manifest: the treatises are several centuries later than the hymns. Some pādas count seven syllables as written where recitation restored a lost syllable; the meter row shows the written count. The treebank's words are unaccented, so the accent comes from the edition alone.

What a specialist could improve. Vowel and consonant values in sanskrit.py; the accented forms in texts/rigveda-1-1/edition.yaml; a rule for the restored syllables, which the meter row does not yet count.

Ancient Greek

Schemes.

Homer and Aristotle open in restored; the New Testament books, Epictetus and Marcus Aurelius open in koine. There is no separate Homeric or archaic scheme. The Iliad and Odyssey manifests relabel restored for their pages: "classical Attic of the 5th century BC with its pitch accent; for Homer, the nearest well-studied stage, not the poet's own (approximate)".

What it rests on. W. S. Allen, Vox Graeca, for restored (module); Randall Buth for koine (module label and the New Testament manifests); the Epictetus notes cite Horrocks, Greek: A History of the Language and its Speakers (2nd edition, 2010), chapter 5. The length of α, ι, υ, which the spelling does not show, comes from quantity tables: grc_quantities.yaml (shared), grc_quantities_iliad.yaml and grc_quantities_marcus.yaml, each entry from the LSJ headword and, for verse, the foot where the hexameter shows it.

The hexameter scanner. greek.scan_hexameter (version 0.1) scores every arrangement of five dactyls or spondees plus a final foot, with a cost for each rule it leans on (epic correption, muta cum liquida, metrical lengthening, an unlisted long α ι υ, synizesis, neglected digamma), and reports each costed choice as a failure class. test_hexameter_scanner_matches_hand_scansion in tests/test_golden.py requires its output to equal every hand scansion in the curated overlays: 62 lines (Odyssey 1.1-10 and Iliad 1.1-52). Not modeled: digamma beyond a short list of stems, lengthening before initial liquids except as a costed option, and synizesis beyond word-final -εω.

Where it is weakest.

What a specialist could improve. Entries in the quantity tables; the digamma stem list and the costs in the scanner; the Koine vowel and consonant tables in greek.py; a scheme for the language of the Homeric poems, which the repository does not yet have.

Biblical Hebrew

Schemes.

What it rests on. Geoffrey Khan, The Tiberian Pronunciation Tradition of Biblical Hebrew (2020) (module). The pointing gives the vowels, dagesh gives gemination and hard stops, and the cantillation accent gives the stressed syllable, so no quantity table is used. Shewa and begadkefat follow "the standard grammars" (module; no grammar is named).

Where it is weakest. The manifest: the vowels are the Masoretes' of about AD 900, so "as first written" reaches the pointing, not the period of composition. Vocal against silent shewa and qamets gadol against qatan are guessed where the pointing is ambiguous. The verse-final silluq is recovered from the sof pasuq.

The divine name. The Leningrad Codex writes YHWH with the vowels of the word read in its place (the qere perpetuum, Gesenius §17c): יְהוָה is read ʾăḏōnāy and יֱהֹוִה ʾĕlōhîm. The sound row gives the reading, with the written word's accent moved onto its stressed syllable, and the surface keeps what is written. After a prefix (לַיהוָה) the reading changes the prefix vowel; that case is not derived and is reported as qere-perpetuum-prefixed. The Shema is the first work where the name occurs. How the name itself was pronounced before it was replaced in reading is not reconstructed.

What a specialist could improve. The shewa and qamets heuristics in hebrew.py, with a named grammar; an earlier reconstruction as a scheme placed before tiberian, if one can be stated per word.

Aramaic. The Aramaic chapters of Daniel carry the same Tiberian pointing, so Daniel 5 is read by this module unchanged, under the language code arc. What differs is the morphology: the OSHB parsing strings are prefixed A, and the verb-stem letters then mean the Aramaic stems (Peal, Pael, Haphel, Hithpeel and the rest), which treebank.py decodes from its own table. Words written one way and read another (Ketiv and Qere) reach the digital text as consonants without vowels, and their sound row is a reading supplied by the reader, which the page's notes say.

The Kaddish. The Kaddish is Jewish Aramaic of the Talmudic period, known from prayer books rather than from a Masoretic manuscript, and its pointing is a modern prayer-book pointing without cantillation. The Tiberian scheme applies the Masoretes' sound values to that pointing: the nearest fully recorded reading, not a reconstruction of the prayer's own time and place, so the work's confidence is low. With no accents to read, the stress is declared in the edition (stress_default: 1, the final syllable, and stress: 2 on the few words stressed on the penult). The prayer-book text writes כָּל without maqaf; the module reads it, after any prefixes, as qamets qatan (kol). The Ashkenazi pronunciation of the prayer (yisgadal v'yiskadash) is not yet a scheme.

Classical Chinese

Schemes.

What it rests on. Old Chinese readings are taken per character from Wiktionary's data modules, each pinned to a revision id, into pipeline/weft/data/lzh_oc_bs.yaml (CC BY-SA 4.0). A character Baxter-Sagart lacks falls back to Zhengzhang Shangfang (2003) and says so. Tang readings are Unihan kTang, which follows Hugh M. Stimson, T'ang Poetic Vocabulary (1976). A character with no reading of its own borrows one from a glyph variant (kZVariant, then kSemanticVariant) and says so.

Where it is weakest.

What a specialist could improve. Entries in lzh_oc_bs.yaml (a reading, or a better fallback for the four characters outside Baxter-Sagart); Tang readings for the missing characters, which need a source other than Unihan; Mandarin choices for polyphonic characters.

Latin

All Latin schemes keep the classical penultimate stress rule (except the French method, below), so every word needs its vowel length from a quantity table: pipeline/weft/data/lat_quantities.yaml, plus per-work tables for the Bayeux Tapestry, Inter caetera, the Ninety-five Theses, the Res Gestae and the Saer de Quincy charters. Entries follow Lewis and Short headword quantities plus inflectional endings, confirmed against the hexameter for Ovid. A word missing from the table is reported as quantity-unknown. weft quantities <work> drafts entries for a new work's unlisted forms: each form's lemma is looked up in Lewis and Short (fetched by weft lexica), and the vowels its headword marks long are carried onto the form's stem. The proposals go to gen/quantities-proposed.yaml for review and reach no page until copied into a table. With the Res Gestae's own table withheld, 90% of the long marks it places agree with that table; it finds about 60% of the long vowels, because a headword shows the stem and not the ending, and a reviewer adds the endings.

Schemes.

IdLabelUsed byWhat it represents
classicalClassical: restored pronunciation of Cicero's and Ovid's Romefirst for Ovid, the Res Gestae and Bede (relabeled for Northumbria about 731); third for Luther; second elsewhereafter W. S. Allen, Vox Latina: c and g hard, v as w, ae as ai, length audible
ecclesiasticalEcclesiastical: Italianate church Latinsecond for Ovid and the Res Gestae; second of three for Luthersoft c and g before front vowels, v as v, ae as e, length heard only in stress
anglo-normanAs first read: Latin in Normandy and Norman England around 1070 (approximate)Bayeux Tapestryc before front vowels ts, g and j dʒ, h silent, u as French u, s between vowels z
danish-medievalAs first read: Latin in Denmark around 1200 (approximate)Saxo, Gesta Danorum (Holder 1886)c before front vowels s, ti before a vowel s (Holder writes ci), g hard, consonantal u and w v, qu kv, h sounded but k in michi, th t, length heard only in stress
anglo-latinAnglo-Latin: as a clerk in England read Latin around 1215 (approximate)Magna Carta; the Saer de Quincy charters; Henry I's Charter of Liberties (relabeled for 1100)soft c ts, soft g and j dʒ, h silent, v as v, s between vowels z
italian-humanistAs first read: Latin in northern Italy in the 1480s (approximate)Pico; Inter caeteraItalian vowels, soft c ch, gn ny, sc sh, ti ts, h silent, s between vowels z
low-countriesAs first read: Latin in the Low Countries around 1500 (approximate)ErasmusDutch vowel values long in open syllables, u as Dutch uu, g a fricative, ch kh, ti ts
tudor-englishAs first read: Latin in England around 1516 (approximate)MoreEnglish long values in stressed open syllables at an earlier stage of the Great Vowel Shift, ti as si
german-humanistAs first read: Latin in Saxony around 1517 (approximate)Luther, Ninety-five ThesesGerman lengthening rule, c before front vowels ts, g hard, qu kv, v as f, final devoicing
as-first-readAs first read: in the author's own country's manner (approximate); each work names its mannerDescartes and Newtonper section dialect: english (1680s English method) or french (1640s French method, final stress, nasal vowels)

Manifests relabel some of these for their page (Erasmus, Pico, Inter caetera, More, Luther, the Res Gestae).

What it rests on.

Where it is weakest.

What a specialist could improve. Quantity-table entries, especially hidden quantities and medieval words keyed by their classical stems; the rule functions in latin.py (_humanist, _german, _norman, _danish, _national) for a named period grammar; an eu diphthong; a separate southern Italian or Roman reading for the papal chancery.

Sahidic Coptic

Schemes.

What it rests on. Peust, Egyptian Phonology (1999), for the letter values; Layton for the reading of a doubled vowel; the letter table published by the Coptic Orthodox Diocese of the Southern United States for bohairic (module). The accent of Greek loanwords and the list of unstressed particles are in pipeline/weft/data/cop_stress.yaml. Lemma and morphology come from UD_Coptic-Scriptorium (CC BY 4.0), which annotates a different digital text of Mark; spelling differences are normalized before the two are compared, and what remains is reported as edition-differs-from-treebank (manifest). Glosses are hand-written against the treebank's lemmas, with Crum's Coptic Dictionary (1939) and Lambdin's Introduction to Sahidic Coptic (1983) as references (manifest).

Where it is weakest. The edition prints no supralinear strokes and no punctuation, so the syllabic vowel is supplied by rule rather than read from the manuscript (manifest). That Greek loanwords keep the Greek accent is an assumption (manifest). The value of ⲩ standing alone in Greek words may already have shifted (key). Each word is phonemized alone.

What a specialist could improve. The stress table and the particle list in cop_stress.yaml; the syllabic-consonant rule and the Greek-letter values in coptic.py; the normalization list used to compare the edition with the treebank's text.

Old Tamil

Schemes.

No syllable is capitalized: the module states that Tamil stress does not distinguish meaning.

What it rests on. "The description of sounds in the Tolkāppiyam and the usual reconstructions of Old Tamil" (module). No modern reconstruction is named. The sound is computed from a hand ISO 15919 transliteration, checked against the script by tamil_to_iso.

Where it is weakest. The date of the text is uncertain (4th to 6th century AD, with earlier estimates), so the reading is "an approximation of an approximate period" (manifest). The values of c, ṟ, ṟṟ, ḻ and the āytam are the least certain. Sandhi across words is written but not modeled.

What a specialist could improve. Consonant allophony and the āytam in tamil.py, with a named reconstruction; the transliterations in texts/tamil-tirukkural/edition.yaml.

Runic inscriptions (Proto-Norse and Old East Norse)

Scheme. as-carved, "As carved: Proto-Norse around 400, Old East Norse around 800". One scheme; each token's dialect (pn or oen) selects the rules. The Proto-Norse z is a buzzing sound; Old Norse ʀ is "between z and r"; v and w are both w; stress always on the first syllable. There is no second scheme.

What it rests on. The module does not cite a source. Sound is derived from each word's scholarly normalization in each runic work's edition.yaml (texts/runes-*/), never from the runes, because the younger futhark has 16 runes for about 30 sounds.

Where it is weakest. Runes do not mark vowel length or many consonant contrasts; the sound inherits every choice in the normalization (manifest). The Gallehus runes survive only in 18th-century drawings.

What a specialist could improve. The normalizations, with a named edition; the values of z and ʀ and of the fricatives in runic.py; a citation for the scheme as a whole.

Old Irish

Scheme. old-irish, "Old Irish of the 9th century (reconstructed, approximate)". One scheme. Each consonant broad or slender by the vowel letters around it; after a vowel, single c t p are g d b and single b d g m are the fricatives β ð ɣ and nasal β̃; ch th ph are x θ f; initial and doubled l n r are the strong sounds L N R; nd is N; stress on the first syllable, or the first after a pretonic preverb; unstressed short vowels ə.

What it rests on. Thurneysen, A Grammar of Old Irish (1946), and Stifter, Sengoídelc (2006) (module). Old Irish spelling does not write the lenition of b d g m, the nasalization of a vowel, the stress of a compound verb, or a division into words, so the sound is read from a normalized spelling in each token's n: long vowels marked, a dot for unwritten lenition (ṁenma), n- for nasalization, · after a pretonic preverb, a hyphen before an enclitic, a diaeresis for hiatus (ru·biäm). The normalization is Weft's, draft. In Pangur Bán every line counts seven syllables, as the metre (deibide) requires; a test holds the count.

Where it is weakest. The quality of a consonant after e, the value of unstressed final vowels, and whether nd and mb were already n and m. Reading every unstressed short vowel as ə flattens distinctions the 9th-century spelling still keeps. The respelling does not show broad and slender.

What a specialist could improve. The normalization of Pangur Bán, token by token; the treatment of unstressed vowels; the lenition of consonants after l and r, which the module applies to b d g m.

Middle Welsh

Scheme. middle-welsh, "Middle Welsh of the 12th and 13th centuries, the language of the manuscripts (approximate)". One scheme. Stress on the last syllable but one; y clear (ɨ) in a final syllable and obscure (ə) before it, and in the proclitics; u a close rounded vowel (ʉ); ll ɬ, ch x, dd ð, f v, ff f; the nasal mutation (ngh, mh, nh) and rh voiceless; w and i consonants before a vowel where modern spelling has them so.

What it rests on. D. Simon Evans, A Grammar of Middle Welsh (1964), and Morris-Jones, A Welsh Grammar (1913) (module). The manuscripts write the sounds inconsistently (u and v for v, d for both d and dd, t for d, gu for gw), so the sound is read from a normalized spelling in each token's n, written in modern Welsh letters with the word's medieval form kept (frein, neithiawr). The normalization is Weft's, draft.

Where it is weakest. The value of u and of the diphthongs au and eu, which changed in this period; vowel length, not modeled; and the poems' composition in an older language than the manuscripts'. The Gododdin may be centuries older than its copy, with the stress then on the final syllable: the scheme reads the manuscript's Welsh, not the poet's, and says so.

What a specialist could improve. The normalizations, token by token; a second scheme for Old Welsh (final stress, the older vowels) for the Gododdin and the oldest poems.

Old English

Scheme. west-saxon, "Late West Saxon, around the year 1000". One scheme. Stress on the first syllable of the root; ge-, be- and, on verbs, a-, for-, of-, on-, to-, un-, ymb- unstressed. Palatal c and g, sc, cg, voicing of f s þ between voiced sounds, and h by position. be and ge before a or o are the root's diphthong (bearn, geard), not the prefix. Cædmon's Hymn relabels the scheme for its two copies: the Northumbrian copy of about 737 is read from its own words, respelled in the edition's n field in the letters the rules read (uard as ward, dryctin as dryhtin).

What it rests on. Consonant rules after Mitchell and Robinson, A Guide to Old English (module). Vowel length comes from the edition's marks (Harrison and Sharp), corrected by pipeline/weft/data/ang_quantities.yaml (values after Clark Hall, A Concise Anglo-Saxon Dictionary), then from the hand-annotated lemma, whose macrons carry over along a shared prefix. ang_prefixes.yaml lists prefixes the rules cannot see without the part of speech.

Where it is weakest. Palatal c and g are inferred from spelling; an edition that dots them would remove the guess (manifest). Forms that diverge from their lemma before the last macron are reported quantity-uncertain. Compound stress relies on the editor's hyphens, and a hyphen also makes the second element begin as a word does (middan-geard has the y sound). The Northumbrian of Cædmon's Hymn is sounded by West Saxon rules: its words and endings are its own, its sound values are the later standard's.

What a specialist could improve. The quantity table (three entries at present) and the prefix list; the palatalization rules in oldenglish.py; a second scheme, since none is offered; an early Northumbrian scheme for Cædmon's Hymn and the other 8th-century records.

Old Norse

Schemes.

What it rests on. E. V. Gordon, An Introduction to Old Norse, and Haugen (module; no Haugen title is given). The normalized spelling marks length, so no quantity table is used.

Where it is weakest. The editions' ö covers both ǫ and ø and is read as ǫ by default (Eddic and Snorri manifests). For Snorri, ǫ and ø were merging about 1220 and long vowels may have begun to change. Grettis saga is read from IcePaHC's modern Icelandic spelling, so its Old Norse reading is reconstructed from modern forms; its notes date the saga to about 1310, later than the scheme.

What a specialist could improve. A per-word distinction of ǫ and ø; rules in norse.py for a later date (about 1300) as a separate scheme; old forms for Grettis saga.

Old East Slavic

Schemes.

What it rests on. The handbooks of Russian historical phonology, which rest on the spelling of dated manuscripts and birchbark letters: Shakhmatov; Kuznetsov; Shevelov, A Prehistory of Slavic (1964); Schenker, The Dawn of Slavic (1995) (module). The text is Weft's transcription of the Laurentian copy of 1377 as printed by Karsky in Полное собрание русских летописей, vol. 1 (1926), checked against the OCR of the scan with declared corrections (manifest). No treebank is used: TOROT covers the chronicle but is licensed NC (manifest).

Where it is weakest. The jers: their loss in weak position is dated to the 12th century, so at 1100 the weak ones were fading, and the scheme treats every written jer alike; the module calls this its weakest point. The value of ѣ varied by region, and г in Kiev was probably a fricative, which is not shown. The manuscript does not mark stress; it is hand-entered in n from the stress of the same words in Russian and Ukrainian and from the accent paradigms of the handbooks (module and manifest). Softening before е, ѣ and и is assumed and not marked. The years, written in letter numerals, are read as numbers; the number words are not reconstructed (manifest).

What a specialist could improve. The stress marks in texts/primary-chronicle-varangians/edition.yaml; the jer rule and a southern variant with fricative г in oldeastslavic.py; the regional value of ѣ.

Persian

Schemes.

What it rests on. The grammarians, early manuscripts, and Dari and Tajik (module). The notes cite Lazard, La langue des plus anciens monuments de la prose persane (1963), for phonology, and Whinfield's notes; the rubāʿī meter follows Elwell-Sutton, The Persian Meters (1976), and Farzaad's division of the line. The sound is computed from a hand transliteration that writes every vowel.

Where it is weakest. Stress is placed by the modern rules, "the least certain part of the reconstruction" (module); the ð rule and the vowel qualities are approximate (manifest). Short vowels are supplied from Whinfield's notes, the meter and the dictionaries and are checked against the script only for its letters. The metre's licenses are a reciter's judgment.

What a specialist could improve. The stress rule and the ð rule in persian.py; the transliterations in texts/persian-rubaiyat/edition.yaml.

Middle Mongolian

Scheme.

What it rests on. The Ming transcription itself: which Chinese characters the transcribers chose for which sounds (aspirated initials for t, č, k, q and unaspirated for d, j, g, b, as Shiratori's preface also notes); the Uyghur-script spelling of later Mongolian; and comparison with modern Mongolian (module). The sound is derived from the romanization row, not from the Chinese characters read aloud. That row is Shiratori's romanization (1943) converted by rule into current conventions and hand-corrected for eighteen words, each correction marked in the word's provenance (module and manifest). Two sensors test it against the Ming spelling: the shoulder marks 舌 and 中 (every r has its mark, every mark its r or q) and vowel harmony (manifest).

Where it is weakest. The whole scheme is a reconstruction (manifest sound_confidence: low). The aspiration contrast is inferred from the Chinese characters; the hiatus written ' may already have been a long vowel by the 14th century, and the scheme does not decide where; first-syllable stress is assumed and not attested for the period (module and manifest). The romanization is a transcription of the Ming spelling, not a reconstruction of the lost Uyghur-script original. The shoulder-mark sensors report disagreements as telemetry rather than failures, because the base text sometimes omits a mark.

What a specialist could improve. The conversion rule and the eighteen n overrides in texts/secret-history-mongols/edition.yaml; the vowel values and the aspiration reading in mongolian.py; evidence for where hiatus had become length.

Japanese

Schemes.

Hyphens divide morae; pitch accent is not marked in either scheme.

What it rests on. Reconstructions that keep づ and ぢ distinct into the 1600s, and the Portuguese missionaries' spellings around 1600 (Rodrigues) for ye (module). The sound is computed from each token's reading in historical kana.

Where it is weakest. The dzu of づ, the ye of え and ゑ, and the ō of medial -afu are each uncertain by the 1680s (manifest and key). The module reads a token written only は as the particle wa; a noun spelled は alone would be misread. Pitch accent is absent.

What a specialist could improve. The three uncertain rules in japanese.py; pitch accent, if a period source supports it; the kana readings in texts/basho-furuike/edition.yaml.

Old French and Franco-Italian

Schemes.

What it rests on. Nyrop, Grammaire historique de la langue française, vol. 1 (1899), and Pope, From Latin to Modern French (1934) (module and notes). The sound comes from pipeline/weft/data/fro_lexicon.yaml, keyed by a hand normalization into central French spelling. For it1300, the spelling of Franco-Italian manuscripts; the module states that no grammarian of the period describes such a reading.

Where it is weakest. When final consonants fell silent before consonants, and whether ch and j were still affricates (manifest; the module calls the first "the weakest point of the scheme"). it1300 is a hypothesis. The speech of the writers (a Venetian and a Pisan) is not modeled.

What a specialist could improve. Lexicon entries; the final-consonant rule in oldfrench.py; evidence for or against the Italian reading.

Italian

Schemes.

What it rests on. pipeline/weft/data/it_lexicon.yaml, following the DOP (Migliorini, Tagliavini, Fiorelli, Dizionario d'ortografia e di pronunzia), checked against the Latin source vowel (module). The notes add Migliorini, Storia della lingua italiana (1960), and Izzo, Tuscan and Etruscan (1972).

Where it is weakest. The vowel lexicon is built on the modern Tuscan standard; learned words are less certain for these dates. Latinizing spellings are read as written, which may be ornament rather than speech. The gorgia is left out as not securely attested this early (module and all three manifests).

What a specialist could improve. Lexicon entries, especially those marked # learned; a separate scheme for 1513 if the evidence supports one; the meter rules for synaeresis and dialefe.

Portuguese

Schemes.

All three read one normalized spelling per word, written by hand in pipeline/weft/data/pt_lexicon.yaml with the stressed e and o marked for quality; a stressed e or o left unmarked is reported.

What it rests on. Fernão de Oliveira, Grammatica da lingoagem portuguesa (1536), and João de Barros, Grammatica da lingua portuguesa (1540), the first native descriptions of the sounds; Paul Teyssier, História da língua portuguesa (1980), and Ivo Castro, Introdução à história do português (2006), for the dating of the changes (module). Stress follows the written rule of modern Portuguese.

Where it is weakest. The final unstressed o read u in 1540, which Teyssier dates to that century but not to a year; the exact values of the apical and dental sibilants; the metre's elisions, which are not modeled since each word is sounded alone (manifest). Camões's own speech is not modeled.

What a specialist could improve. The quality marks in the lexicon; the unstressed-vowel rules of the 1540 scheme in portuguese.py; a rule for synalepha in the decasyllable.

Spanish

Schemes.

What it rests on. Nebrija, Gramática de la lengua castellana (1492) and Reglas de orthographía (1517), and his Vocabulario for aspirated h; Lapesa, Historia de la lengua española; Penny, A History of the Spanish Language (module). The sound comes from pipeline/weft/data/es_lexicon.yaml, a normalized period spelling per word.

Where it is weakest. The value of h from Latin f, the b and v distinction, and the quality of the sibilants (manifest). How far ts and dz had lost their stop element by 1492 is uncertain (module). Columbus's own Genoese and Portuguese-colored speech is not modeled.

What a specialist could improve. Lexicon entries; the b and v assignments word by word.

Dutch

Schemes.

What it rests on. The Twe-spraack vande Nederduitsche letterkunst (1584); de Heuiter, Nederduitse orthographie (1581); Schönfeld, Historische grammatica van het Nederlands, revised by van Loey (1959); de Vooys, Geschiedenis van de Nederlandse taal (1952) (module). Syllables and stress come from pipeline/weft/data/nl_lexicon.yaml.

Where it is weakest. Whether ij was a diphthong or still [iː] in Amsterdam in 1596; the n of -en; the stress of French and Latin loans (manifest). The sound is computed from modern spelling, so the sharp-long and soft-long ee and oo are not shown.

What a specialist could improve. Lexicon entries; the two long ee and oo, if a per-word source exists.

French

Schemes.

No syllable is capitalized: the module states that French has no distinctive word stress.

What it rests on. For fr1791: Féraud, Dictionnaire critique de la langue française (1787-88), and the grammarians of the decade as collected by Thurot (module). For m1580: Meigret (1542, 1550), Peletier du Mans (1550), Ramus (1562, 1572), Bèze (1584) and Henri Estienne (1578), as collected by Thurot (1881-1883); Palsgrave (1530) on final consonants (module). For fr1885 the evidence is direct, not reconstructed from rhymes or spelling: Passy, Les sons du français (1887), the transcriptions of Le Maître phonétique (from 1886), and Littré's Dictionnaire de la langue française (1863-1872), which gives a pronunciation for each word (module). The sound comes from pipeline/weft/data/fr_lexicon.yaml; where the lexicon has no fr1885 form, the word is taken from its modern form with the length rule applied.

Where it is weakest. m1580: the value of oi, the r of -er infinitives, and the strength of the e caduc (manifest). Montaigne's Gascon-colored speech is not modeled. fr1885: vowel length is shown for the word said alone and was reduced inside a phrase; the optional liaisons of formal reading are marked by judgment, following Littré where he gives a note; the scheme models the Paris norm of the language the Act was written in, not the accents of the delegates, who were not French (manifest).

What a specialist could improve. Lexicon entries, including fr1885 forms from Littré; the liaison exceptions (nz) and formal liaisons (zf) in the editions; the length rule in french.py.

German

Schemes.

What it rests on. For northern: Viëtor, Die Aussprache des Schriftdeutschen (1885); Siebs, Deutsche Bühnenaussprache (1898); von Polenz, Deutsche Sprachgeschichte, vols. 2 and 3. For ecg1545: Luther's own spelling and the Saxon chancery usage; Ickelsamer (1527, about 1534); Frangk, Orthographia (1531); von Polenz, vol. 1; Besch, Luther und die deutsche Sprache (2014) (module). Syllables and stress come from pipeline/weft/data/de_lexicon.yaml. Kant's Latin phrases are given in the German school pronunciation of Latin (manifest).

Where it is weakest. northern: postvocalic r and long ä (manifests); no description of either author's speech; regional features not modeled. ecg1545: the diphthong values, the extent of the p/b and t/d merger, the value of w, and whether unrounding of ü and ö reached a reading of Scripture (manifest). Dauid with f is a judgment (manifest).

What a specialist could improve. Lexicon entries, including old vowel lengths where they differ from modern ones; the lenis rule and w in german.py.

Swahili

Schemes.

What it rests on. Steere's own account of his spelling in Swahili Tales (1870), preface, p. xiv ("the vowels are pronounced as in Italian, the consonants as in English, and ... there is always an accent on the last syllable but one"), and the letter values in A Handbook of the Swahili Language as Spoken at Zanzibar, third edition (1884), pp. 8-15 (module). The text is Weft's transcription from the page scans, checked against the Wikisource transcription and the OCR, with the OCR's misreadings declared token by token (manifest).

Where it is weakest. The Handbook used is the third edition, revised by Madan after Steere's death, and no recording exists (manifest). Not modeled: the aspirated p, t and k Steere hears where a nasal has been lost; the implosive b, d and g that later descriptions report; tone and intonation (module). The long consonants of Arabic loans are uncertain, since the Handbook notes a tendency to drop one of them (module and manifest).

What a specialist could improve. The syllabic-nasal rule and the a + e coalescence in swahili.py; the modern forms in texts/swahili-tales-steere/edition.yaml; aspiration where a nasal was lost.

Cherokee

Schemes.

Both schemes give one syllable per sign. A lone s (Ꮝ) closes the syllable before it, or begins the first syllable of a word. A romanization row under the text is generated from the signs in the transliteration that descends from Worcester's chart and that the Unicode character names use. The module reads its sign table from those names, so no sign value is typed by hand (module).

What it rests on. Worcester's answer to Rafinesque's questions, Cherokee Phoenix, vol. 1, no. 23 (6 August 1828), p. 2, as transcribed by Western Carolina University's Hunter Library. Of the gi in the nation's name he writes "G. as pronounced by Germans represents it best", and of five g-signs that "the sound of German G is most common". Of six signs that write English D, "the consonant sound of the three first of which is always, and that of the three latter often, better represented by D. than by T."; the letter does not name the six, and the module takes them to be da, de, di and do, du, dv from the recitation order. v has "nearly the sound of un French". English J, Ch and soft G in proper names are written with the ts-row (Tsa-tsi, George), and the press's dialect has l where "one dialect" has r. Everything else in the first scheme (ka and the t-row aspirated, hna voiceless, the laterals, w and y) and all of the modern scheme follow present descriptions of the syllabary's values (module).

Where it is weakest. The syllabary does not write tone, pitch accent, vowel length, a syllable-final h or glottal stop, or whether a written vowel is said. A vowel may be written only to carry a consonant cluster (ᏧᎾᏍᏗ is said with no vowel after n). The sound row therefore says every written vowel, carries no accent, and gives no length or pitch. It is a reading of the signs, not a record of a speaker. Worcester notes that many syllables are "sometimes aspirated" (hwi, hli, and no "as if written hno"); the row does not model this. The ts-row is read as ts in the first scheme, as Worcester spells it, but his letter does not settle between ts and the j of jam, and present descriptions give both by dialect. The obsolete sign Ᏽ is read as mv, after its Unicode name. Worcester's letter describes two aspirated nasal syllables without naming the signs, and Walker and Sarbaugh (1993), as cited by Wikipedia, read him as placing this one near hv; the press cast no type for it (module).

The first work. The preamble of the Constitution of 1827, from the bilingual printing of 1828, transcribed by hand and by Tesseract and settled on a scan of about 180 pixels to the inch. Two habits of the 1828 type matter for any reader of it, by eye or machine: Ꮩ (do) is cast as a Λ, an A without its crossbar, and Ꮅ (li) as a plain F; Tesseract's model reads them as Ꭺ (go) and Ꮲ (tlv). Where the scan alone did not settle a sign, the word was looked for in the 1860 New Testament and the dictionaries at cherokeedictionary.net, and two signs are left uncertain on the page. The glosses are drafted against Feeling's dictionary, the New Testament and Wiktionary, and are held as a draft until a Cherokee speaker has reviewed them (work).

Reading the type, measured. Article I (77 printed lines, 1,280 signs) was read by eye before any OCR of its pages, the reading frozen with a hash, and then collated with Tesseract and settled on the scan. The hand reading differed from the settled text at 11 signs (0.9 percent), Tesseract's witness, as pinned in the edition, at 242 (18.9 percent, by edit distance). The hand figure is a lower bound: the places where the two readings disagree were settled, and a sign both misread shows only when something else catches it. One was caught while the article was being glossed: both readings had Ꭶ (ga) as the second sign of a word the dictionary has with Ꮥ (de), ᏭᏕᎵᎬ, "west", and the scan, looked at again, is consistent with Ꮥ. The hand reading was blind to the OCR, not to the lexicon, which chose between look-alike forms. Seven of the eleven hand errors are one confusion: the type has two U-shaped signs with a bar, a plain one, Ꮜ (sa), as in ᎠᏂᎫᏌ "Creeks", and one whose left arm curls. The hand reading took the curled sign for Ꮜ in all seven places; Tesseract read Ꮰ (tlo), and the collation settled on Ꮰ by the curl. Two words later bore it out by their stems: ᏚᎾᏠᎯᏍᏓᏅ, "they agreed on", beside "the Treaties concluded", shares its stem with ᎧᏃᎮᏛᏓᏠᎯᏍᏛ, "treaty", in the Cherokee Nation's word list and with ᎧᏃᎮᏛ ᏚᎾᏠᎯᏍᏛ, "covenant", in the 1860 New Testament (Galatians 3:15), and ᏚᏠᏏᎸ, where two rivers join, shares its stem with ᏕᎬᏩᏠᏏᎴᎢ, "they came out to meet him" (Matthew 8:34). With Ꮜ in its place, neither stem was found. The other five places where the sign stands are in words not yet found. The same reading found the other pairs a reader of this type must separate: a G with a straight spur is Ꮯ (tli, as in ᎠᏟᎶᏛ), with a curled spur Ꮳ (tsa, as in ᏣᎳᎩ), with a spiral inside Ꮆ (lo); Ꮥ (de) has its bar at the upper right, Ꭶ (ga) through the middle; Ꮲ (tlv) is a P with a closed bowl, and Ꮅ (li) is an F, flat or with its arm curled down; Ꮔ (nu) has a loop at the upper left, Ꮈ (lv) a tick at the middle left.

What a specialist could improve. Where written vowels go unsaid; the aspiration Worcester describes; tone, from a dictionary that marks it (Feeling, Cherokee-English Dictionary, 1975) word by word in the edition. Corrections from speakers come first (see About Weft).

Esperanto

The first invented language in the public library, and the second (after Tolkien's, private) whose author wrote down its pronunciation. Zamenhof's Unua Libro (1887) gives the alphabet with a sound for each letter and the rule that the accent falls on the last syllable but one; the Fundamento (1905) fixed both (rules 9 and 10). So the one scheme is the author's rule, not a reconstruction:

What the rule leaves open and Weft decides: the exact vowel qualities (Zamenhof gave them by comparison with other languages), and where a consonant cluster divides between syllables (es.tas, pa.tro: s + stop divides, obstruent + liquid holds). No lexicon is needed; the sound follows from the letters. The module also reads the grammar: the 1887 print divides each word into its parts (Patr'o ni'a), and esperanto.analyze turns the division into lemma and Universal Dependencies features by the Fundamento's rules, which the overlay declares as an automatic analysis checked by hand. The 1887 spellings that differ from later usage (Si for Ŝi, tranquil- for trankvil-) are kept in t with the regular form in n.

Quenya and Sindarin (Tolkien)

No public work is in either language: Tolkien's texts are in copyright until 2043, so Namárië and the hymn to Elbereth exist only as private works built from a reader's own copy (private/README.md). The module is public because it states pronunciation rules and holds no text.

Scheme.

Script row (Tengwar). Each word also carries a Tengwar row, set in Alcarin Tengwar (SIL OFL 1.1, bundled in site/fonts, loaded only on pages with Tengwar). Quenya follows the classical mode of Appendix E as Tolkien wrote Namárië in Tengwar on p. 65 of The Road Goes Ever On (3rd ed.): each vowel sign over the consonant before it, the long carrier for a long vowel but a doubled accent for long é, yanta and úre for the second part of a diphthong, rómen before a vowel and óre elsewhere. Ten words are pinned in the tests as read off that page; the rest await a reading against it (private/tolkien-namarie/review/, weft#196). An edition token may carry tengwar: to give Tolkien's own spelling where it departs from the mode. Sindarin follows Appendix E's general use (the vowel sign over the consonant after it) and is Weft's transcription: Tolkien left no Tengwar of these Sindarin texts.

Pauses (Namárië). Tolkien's handwritten copy of Namárië (The Road Goes Ever On, p. 76) marks "the pauses in the chant in ascending order of time-length and tone-fall" with | || |||. They are kept as a pause field on the word each follows (private/tolkien-namarie/curated/pauses.yaml, sixteen marks) and shown after that word's respelling. The same page divides each line into feet with thin hairlines; a hairline that falls between words is told from a pause by the iambic scansion. The hairlines are kept as each line's metre string, the line's own text with | where Tolkien drew one, and only there: he did not mark every boundary (none after the opening trochee of line 3, one only in line 17). Thirteen lines are recorded; lines 10, 12 and 14 and the left half of line 4 show only faint traces in the scan and await a reading against the page (weft#200).

What it rests on. Two accounts by the author. Appendix E, part I, "Pronunciation of Words and Names", in The Lord of the Rings (1955), gives the letter values and the stress rule. The "Notes and Translations" in The Road Goes Ever On (1967; read from the 3rd edition, HarperCollins 2002, pp. 66-72) add what Tolkien thought Appendix E had given "not perhaps with great clarity": for Quenya, the tone of the main stress, the secondary stresses, and Namárië printed with every major and minor stress marked; for Sindarin, the short vowel qualities, the sound of ae, the long diphthongs, and initial ng. Where the two differ in detail, the later notes are followed. This is the one language in the library whose author wrote down its pronunciation, so the scheme is a statement of his rules rather than a reconstruction. The regression tests are the twelve stress examples of Appendix E and the marks on Namárië: the module's main stress agrees with every major mark, and its main and secondary stresses together agree with the marked syllables of 36 of 37 words of more than one syllable. The exception is laurië in line 1, where the rule gives the last syllable a light stress that Tolkien describes as available to the metre and the line does not use. Lemma and gloss in the private works follow Tolkien's own word glosses through Eldamo's citations to the page and line; Eldamo's neo-Eldarin forms (fan reconstructions) are never used.

Where it is weakest. Appendix E gives letter values "approximately" and only by English keywords, and the 1967 notes add more keywords (sick, bed, hot, foot) that "may" render the Sindarin short vowels; the IPA stays broad rather than reading lax English vowels into Sindarin. The quality of Sindarin oe is not described (the respelling reads it oi, which Appendix E allows); ae is "very similar to" the vowel of high (1967). How much longer the first vowel of a long diphthong is, Tolkien says only "somewhat". Written th in Quenya is read θ, though Tolkien says the sound had become s in speech. In hyphenated compounds the 1967 marks vary (aíre-tári-lírinèn has three major stresses, ùndu-láve a minor first part); the module gives each part its own main stress. Secondary stress in Sindarin is not described and not shown. Tolkien's 1952 recordings are not used: they differ from the printed text in places and are a performance, not a rule.

What a specialist could improve. The digraph list and the syllable split in elvish.py; whether ly, ny, ry count as one consonant or two for stress (Tolkien's 1967 marks on ómaryo and Calaciryo support two, as the module has it); the Sindarin long-vowel qualities; whether the light final stress should be shown only where a line's metre uses it (laurië).

Open inconsistencies

Known inconsistencies between the modules, the manifests and other documentation, listed so a reviewer can weigh them. Several found while this page was written have been corrected (the Koine and Tang labels on later or earlier texts, the Akkadian and runic language names, the scanner notes, the README sample respelling, the Grettis saga scheme order, a Spanish key example).