저장소 정보는 아직 수집하지 않았습니다.
아직 감지된 업데이트가 없습니다. GitHub 릴리스가 나오면 여기에 보입니다.
최근 완료된 8시간 회차 사이 GitHub 스타 증가량 순
면접·시험 도우미, AI가 실시간으로 풀이 아이디어를 제공
ESP32‑S3 N16R8 벤치 가이드: 실시간 핀아웃, 강의, 시뮬레이터 및 벤치 로그
보편적 교육을 위한 오프라인 앱, Kolibri Learning Platform
학술 논문의 참고문헌을 검증하는 도구
MCP로 AI 어시스턴트를 Canvas LMS와 연결, 학생·교사를 위한 최대 103가지 도구 제공
메이커가 쓴 소개는 위의 한 줄이 전부입니다. 더 자세한 내용은 저장소 README에서 읽을 수 있습니다. 저장소 README ↗
README 전체를 불러오는 중입니다.
README 전문 보기 ↗greek-phonology helps you work with how Ancient Greek sounds. It turns Greek text into IPA under four pronunciation schemes (Modern, Koine, restored Attic, Erasmian), and guesses which scheme a speaker used from their IPA. It also splits words into syllables, checks and places accents, and looks words up in a built-in New Testament and Septuagint word list. It is written in MoonBit and ships as WebAssembly and JavaScript, with no server and no dependencies. Try it in the playground.
npm install greek-phonology
JavaScript (ESM, Node 18+):
import { ipa, detect, call, loadLexicon, lookup } from "greek-phonology";
ipa("ἐν ἀρχῇ ἦν ὁ λόγος", "koine"); // "ɛ n | a r kʰ e | e n | o | l o ɣ o s"
detect("e n | a r x i").probs; // { modern: …, koine: …, attic: …, erasmian: … }
call("accentuate_verb", "λυομεν"); // "λύομεν"
loadLexicon(); lookup("λογος").form; // "λόγος"
WebAssembly (browser, wasmtime, any host). The package ships greek-phonology.wasm (also exported as wasmPath). Strings go in and out one UTF-8 byte at a time, so you don't need a shared allocator. See docs/app.js for a 20-line wrapper:
import { readFileSync } from "node:fs";
import { wasmPath } from "greek-phonology";
const { instance } = await WebAssembly.instantiate(readFileSync(wasmPath), { spectest: { print_char() {} } });
// in_clear(); in_push(byte)...; in_push(0); in_push(byte)...; n = run(cmd); out_get(0..n-1)
// cmd 0 detect, 1 ipa, 2 word, 3 load word list, 4 lookup, 5 call(name, arg), 6 g2p_json
curl -fsSL https://cli.moonbitlang.com/install/unix.sh | bash
cd moonbit && ./tools/build_pages.sh # writes docs/greek-phonology.{wasm,js} and the word list
A release wasm build is 178,373 bytes (66,993 bytes gzipped) with moon 0.1.20261023. docs/greek-phonology.wasm is the copy the playground serves and can lag that build. moonbit/tools/check_artifacts.sh compares them; the pre-commit hook does not run it. The word list (greek-phonology-lexicon.bin, 370,483 bytes, 167,674 bytes gzipped) is optional. Load it once (loadLexicon()) to turn on lookup and to restore accents and breathings on bare input before G2P.
Every function below can be called through call(name, arg, arg2). The most common ones also have their own JS functions.
Pronunciation
ipa(text, scheme): Greek to IPA. Phones are separated by spaces and words by |. ipa("θεός", "attic") → tʰ e o s. Attic follows Allen's 4th-century target (ου [uː], length shown where the spelling marks it, geminates as Cː), Erasmian χ is [x], and Modern voices σ to [z] before β γ δ μ ν λ ρ. Koine word-final αυ/ευ/ηυ is [v]; before a voiceless consonant it is [f], and medial before a vowel or voiced consonant it stays [β]. See docs/RESEARCH.md.g2p(text, scheme, flags?): the same phones, plus one unit per grapheme. Concatenating units[].t gives the typed word back. restored is true when a phone used a mark the user did not type. Flags are a comma list: alts adds tolerated hearings, ctx voices word-final αυ/ευ/ηυ from the next word. Both are off by default, and neither changes ipa(). Wasm command 6. g2p("αι", "modern") → {v:1, scheme:"modern", lib:"0.3.2", words:[{text:"αι", phones:["e"], units:[{t:"αι", k:"diph", ph:["e"], …}]}]}.phoneDistance(target, heard, scheme?, goal?): graded distance between one target phone and one heard phone, for any recognizer's IPA. goal is strict, gracious (default), or lenient. It changes tier and sim only; d, note, and advice stay put. Tiers are accurate, tolerated, and not-a-match. Unknown phones return not-a-match and do not throw. This call stays on the hand-weighted feature table. Sourced rows, including the policy rows, are default off; see docs/DISTANCE_TABLE.md. phoneDistanceIn is the opt-in. It takes left,right,right2,place,posterior when a row needs neighbors or a confidence gate. With no posterior, a gated row is not a match.phoneCredit(target, scheme, goal, posterior): credit for a comma list of phone=weight pairs. An exact hit outranks a near miss, and nothing outside the tolerated tier scores.hearingCredit(target, posterior, scheme?, goal?): the same numbers as phoneCredit, plus the note for the best alternate. posterior may be the list or a candidate array (, , or ).Syllables and accent
syllabify(word): syllabify("ἄνθρωπος") → [["α","ν"],["θ","ρ","ω"],["π","ο","ς"]]word(word): syllables, the written accent, whether it obeys the law of limitation, and the recessive accent.accentuate_phrase(text) / check_phrase_accents(text): accents in context (enclitics, proclitics, grave, elision). accentuate_phrase("ἄνθρωπος τις ἦλθεν") → ἄνθρωπός τις ἦλθενaccentuate_verb(word): places the recessive accent. accentuate_verb("δουλος") → δοῦλοςaccentuate_last_vowel(s): accentuate_last_vowel("χα") → χάadd_acute(s) / add_circumflex(s): add_circumflex("α") → ᾶremove_acutes(s), remove_graves(s), remove_circumflexes(s), contains_circumflex(s), without_accents(s): without_accents("ἄνθρωπος") → ανθρωποςis_long_syllable(s): is_long_syllable("τη") → trueForms
amalgamate(s): consonant + σ. amalgamate("γραφσω") → γραψωcontract(s): vowel contraction (Black's table). contract("ποιεω") → ποιωverb_stem(lexical_form): verb_stem("λύω") → λύaugmentize(s): augmentize("λυον") → ἐλυονconjugate(stem, paradigm): six forms. conjugate("λυ", "future-active-indicative") → λύσω λύσεις λύσει …Text
remove_punctuation(s), remove_numbers(s), extract_consonant_clusters(text): extract_consonant_clusters("στρατος κτισις") → ["στρ","κτ"]moon prove machine-checks these. With Z3 4.13.0 the verifier reports 227 proved goals (227 valid, none invalid or timed out). The count depends on the MoonBit and Z3 versions. Run python3 moonbit/tools/coverage.py after moon prove for the local count; see docs/VERIFICATION.md. The library's real code paths call these proved functions:
accentuate_verb and conjugate always lands where the rule allows.ipa (all four schemes): the rules are finite tables plus a proved step function over code points. Every code point is consumed (each step advances 1 or 2 and the steps end exactly at the word's end), the step is a pure function, and every phone it emits is in that scheme's sound inventory. Koine word-final αυ/ευ/ηυ emits [v]; that postcondition sits inside the existing g_diph_seq check and does not add a goal. Attic vowel length (aː iː yː from a circumflex, macron or iota subscript) goes through a proved helper that stays in the Attic inventory; reading the marks and writing geminates as Cː is tested, not proven. Grapheme-unit tiling and the distance table are tested, not proved.sourced_distance_buckets); the default alignment path does not feed those rows in. An equal-cost tie takes the identical string, then the higher posterior. Near-miss credit is on only when the switch is on and the goal is not exact. A supplied substitution price replaces the table price.cap * sim / 1000. The display rank stays in 0 through 5. Two-call monotonicity lemmas timed out and are tested instead.contract: a table plus a proved step function over code points, with the [ει]/[ου] markers as single tokens. The output is never longer than the input, and every output token is a valid letter, mark or marker.Tested, not proven:
Double. This prover has no useful theory of Double, so monotonicity of those costs is not a goal. The same limit applies to proving that 1,000,000 steps always cover a matrix: each step's rank falls, which rules out a cycle, but a cap of (heard + 1) * (phones + 1) is too small once a restart revisits a cell, and "the cap is at least the matrix size" is false once the constant saturates.without_accents and syllabify give the same syllables in either order. The exception is diaeresis, which marks a syllable break, so removing it rightly changes the split.Parity tests compare the library with the original TypeScript (its outputs are frozen in moonbit/tests/kl_golden.json.gz; run node moonbit/tools/kl_parity.cjs).
moonbit/kernel/ — proved core (moon prove), split across syllable.mbt, law.mbt, contract.mbt, phrase.mbt, detect_arith.mbt, g2p_core.mbt, align.mbt, and metric.mbt. The prover still uses one logic file, spec.mbtp.moonbit/lib/ — syllables (syllable.mbt), accents (accent.mbt), G2P (g2p.mbt, g2p_units.mbt), the opt-in distance table (phondist.mbt, phondist_adapt.mbt), restart-tolerant alignment (align.mbt, align_json.mbt), morphology (morph.mbt), marks (marks.mbt), call_json (call.mbt), the lexicon, and JSON. Reusable logic lives here, not in the playground.moonbit/api/ — WebAssembly byte ABI and the JavaScript exports.moonbit/tests/ — property tests, the ipa() golden snapshot, and the cross-word snapshot.moonbit/tools/ — generators, the Pages build, and the golden and verify scripts.moonbit/data/ — the lexicon blob and the word counts used to build the detector tables.docs/ — the GitHub Pages playground (index.html, app.js) plus committed wasm, JS, and lexicon copies, and RESEARCH.md.notes/ — working notes. The build does not read them.packages/npm/ — npm package source (package.json, index.js, index.d.ts). moonbit/tools/build_npm.sh copies the release JS, wasm, and lexicon in. Nothing here publishes to npm.moonbit/moon.mod and the npm package greek-phonology use the same version number. Published npm releases are 0.1.0 (2026-10-22 18:08Z) and 0.2.0 (2026-10-22 20:05Z). The working tree is 0.3.2 and has not been published; see CHANGELOG.md. This repository has no git tags. Creating tags is a separate push and is not done from documentation changes. The proved-goal count is not a version: moon prove on your machine prints whatever MoonBit and Z3 you installed.
sh scripts/setup-hooks.sh
The hook runs moon check, moon test (including the golden snapshot), an additive .mbti check, and moon prove when z3 is on PATH. Details are in CONTRIBUTING.md. Alignment, the distance table, credit, and schemes are summarized in docs/ARCHITECTURE.md.
The NT forms come from the SBL Greek New Testament (CC BY 4.0). The Septuagint forms are Rahlfs 1935 (public domain), via eliranwong/LXX-Rahlfs-1935. See moonbit/data/README.md.
MIT for the code. The word list carries the attributions above.
phone=weight{phone, p}{phone, probability}[phone, probability]unitHearingCredit(unit, posterior, scheme?, goal?, at?): the same, for one grapheme unit from g2p. A curated alternate note (for example Koine word-final β beside v) replaces the generic note when that hearing is the best near miss. Omit at to keep the phone in the unit with the highest combined credit; pass a phone index to score only that phone.creditLevel(combined, scale?): a display level for a combined credit from 0 to 1. The levels are none, close tint, light, medium, clear, and deepest. The default cutoffs are 8, 30, 60, 85, and 95 percent, exported as CREDIT_LEVEL_CLOSE, CREDIT_LEVEL_LIGHT, CREDIT_LEVEL_MEDIUM, CREDIT_LEVEL_CLEAR, and CREDIT_LEVEL_DEEPEST. Pass scale ({close, light, medium, clear, deepest}, each a percent) to override any of them.creditBands(scale?): those six bands as {level, from, until} percents. A page paints from this table instead of copying the cutoffs.schemeId(name): the short id ipa expects. The playground page uses koine, attic, erasmian, and modern. Longer ids such as attic-reconstructed, grc-koine-reconstructed, grc-attic-restored, grc-modern-style, and grc-erasmian-us map to the same four. An unknown name returns "". ipa itself still accepts only the short ids.alignPhonesJson(spec): align one recognizer take to target phone units. spec is a JSON string or the same object: log-posteriors (lp, T, V, spf), vocab, lane, and optional params. Pass lp as a Float32Array, Float64Array, number array, or base64 of little-endian float32, or call alignSetPost(lp, T, V, spf) and then { postRef: true }, so the matrix is not a JSON number list. lanes scores every lane from one soft decode and returns { tokens, takes }; each lane may set scheme or params. windows plus totalSec stitches overlapping matrices first; without a lane or lanes that call returns the stitched matrix. Each word is credited from its best coherent attempt. A restart may go back three words, a skip costs one per word, and stray tokens are insertions. Substitution cost is phone_distance at the gracious goal: 0 accurate, 0.5 tolerated, 1 otherwise. e/i and r/l cost 1 unless params.subCost names the pair, for example [["e","i",0.5],["r","l",0.5]] (either order). The same string stays 0. params.goal changes near-miss credit on score and overall only, not the path. params.nearMiss: false or goal: "exact" scores only the posterior share on the target or a lane alternate. When two heard tokens match one target at the same cost, the identical string wins, then the higher posterior on that target. Both numbers are 0–100; the JSON also has credit, level, overallCredit, overallLevel, and attemptCredit ({w, att, credited, score} for every attempt). A side over 20,000 phones or a matrix over 250,000 cells returns no pairs. Call name align_phones. ipa() does not call it.optionalPhones(phones): default abstain flags. Rough breathing h is optional. pʰ is not. Call name optional_phones.distTable(scheme): the learner-goal thresholds, the scheme notes, and the curated grapheme alternates (Koine word-final [f] for αυ/ευ/ηυ, with a "more accurate" note; Erasmian υι beside the classroom [w]).detect(ipa, greek_text?): scheme probabilities, plus the evidence behind them. It is more reliable when you pass the Greek text. It allows for typical phone-recognizer mistakes, using a simple prior (the expected phone 80% of the time, any other phone uniformly otherwise); no audio data is used. Phone-bigram tables come from SBLGNT and Rahlfs LXX word counts. The distance table does not feed the detector.lookup(word): finds a word in the NT/LXX list, ignoring case or accents if needed. lookup("λογος") → {form: "λόγος", nt: true, lxx: true, ...}accentuate_phrase): every per-word decision comes from one proved function. A grave appears only when an ordinary word follows. An extra acute appears only before an enclitic and only on a proparoxytone or properispomenon. A word never gets more than two accents. A two-syllable enclitic keeps its accent only after a paroxytone. Undoing the decision (as lexical_phrase does) gives back every ordinary word's dictionary accent.detect probabilities: every probability is at least 0 and they add up to exactly 1. Internally they are integer shares of 2^30, and the rounding remainder goes to the provably strongest scheme. Evidence hits never exceed totals.amalgamate: the output is never longer than the input, and each step either copies a letter or merges a stop+σ into ψ, ξ or σ.