vec4ipa — IPA/extIPA ⇄ 16‑D articulatory vectors

vec4ipa is a suite of three command-line tools that convert between IPA / extIPA strings (including complex combining marks, ligatures, tone letters, Chinese tone classes, and clinical-phonetics symbols) and the 16‑dimensional articulatory vectors defined in this repository (docs/SPEC-NEXT.md — the current scheme; docs/SPEC.md and IPA_VECTORS.md document the legacy v8 scheme, kept for reference).

All three tools share one core (src/ipa2vec_core.h):

tool	direction	purpose
ipa2vec	IPA → vectors	parse IPA/extIPA strings into 16‑D vectors (2‑layer IR, JSON)
vec2ipa	vectors → IPA	nearest segment + modifier fit, distance
vec4ipa	both + inventory	full table, modules, query, stats, weights, both directions

Build

Requirements: C11 compiler (gcc recommended), python3 (only to regenerate generated headers).

make            # builds ./ipa2vec ./vec2ipa ./vec4ipa
                # (Windows: wmain + -municode for UTF-8 argv;
                #  plus ./vec4ipa_ui, the Win32 GUI wrapper in ui/)
make gen        # regenerate src/vectors.h and src/readme_embed.h

The WinUI 3 front-end (ui/winui/, the modern workbench) is built separately on Windows:

ui/winui/build.bat   # builds ipa2vec_core.dll + publishes dist\vec4ipa_ui.exe

The binaries embed the full table; they do not read any file at runtime.

GUI workbench (Windows)

The modern front-end is the WinUI 3 workbench (ui/winui/ → vec4ipa_ui.exe, built by ui/winui/build.bat; the output is a self-contained dist\ folder — see ui/README.md). The UI is Fluent-designed: a Mica window backdrop, in-app acrylic popovers and pickers, theme-aware surfaces:

• IPA → vectors — type IPA or click the grouped soft keyboard (Consonants, Non-pulmonic, Vowels, Diacritics, Letters, Tones, Recent) with a live filter; output as vectors, two-layer IR or JSON
• Vector → IPA — 16-D input (multi-line batch) with a long-press narrowness picker (levels 0–4 by name, magnetic cursor) and reverse fit
• Loop — IPA → vectors → reverse fit with the tone annotation round-tripped (tone=(…) groups survive the round-trip)
• Format picker — long-press the vectors/query/layers/JSON button for an animated slide-to-select list
• Compare symbols, weighted distance, history, examples, open-file batch convert, views (table/modules/stats/weights), CSV/IR export, bundled CLI tools, English/中文, system/light/dark themes

A legacy Win32 wrapper (ui/vec4ipa_ui.c, built by make ui/vec4ipa_ui.exe) is kept for reference: it embeds the three CLI tools as resources and can export them to a folder of your choice (File > Export tools…), so a single shipped exe carries the whole tool suite. It uses an Acrylic-style translucent backdrop (DWMSBT_TRANSIENTWINDOW, blur-behind on Windows 10, none on older systems).

Fonts

The GUI text uses Gentium Book Plus and the application icon (ui/assets/final_ink.svg) renders its arrow glyph with NewComputerModern10 — both SIL OFL 1.1 licensed. The files are shipped in ui/fonts/ (GentiumBookPlus-Regular.ttf, NewCM10-Bold.otf) together with the license text (OFL.txt). The GUI loads them privately (AddFontResourceEx FR_PRIVATE) from fonts/ next to the executable and falls back to system-installed faces when the folder is absent; no font installation is required.

Usage

ipa2vec <STRING>            parse each segment -> vectors
ipa2vec -L <STRING>         two-layer tier decomposition, then rebuild IPA (inverse demo; alias --ir)
ipa2vec -j <STRING>         JSON output
ipa2vec -A <IPA1> <IPA2>    sequence (syllable/word) alignment distance
ipa2vec -M FILE             load metric.json weights/lambda at runtime (or --metric FILE)
ipa2vec -v                  version

vec2ipa <V0,...,V15>        nearest segment + modifier fit -> IPA
vec2ipa -n <V0,...,V15>     nearest base segment only
vec2ipa -d <A> <B>          weighted distance (Mahalanobis + airstream penalty λ)
vec2ipa -A <IPA1> <IPA2>    sequence (syllable/word) alignment distance
vec2ipa -N <LEVEL>          transcription narrowness (broadest|broad|medium|narrow|narrowest, or 0-4; alias --width; default narrow)
vec2ipa -M FILE             load metric.json weights/lambda at runtime (or --metric FILE)
vec2ipa -S <CLASS>          reverse symbol class (standard|extipa|sinologist|all; aliases std, ext, school, sino; alias --charset; repeatable;
                              standard IPA incl. ɚ ɞ ɝ ꞎ ᶑ never gated; extipa = ʬ ʭ ʩ;
                              sinologist = ᴇ ȶ ȡ ȵ ȴ)

vec4ipa -t                  full base table (main + extIPA bases)
vec4ipa -m                  regional modules and their symbols
vec4ipa -q <SYM>            query a symbol, or parse a base + modifier string into a natural-language description
vec4ipa -s                  statistics
vec4ipa -w                  metric weights / lambda
vec4ipa <STRING>            forward: IPA -> vectors
vec4ipa -j <STRING>         forward: IPA -> vectors, JSON
vec4ipa -L <STRING>         forward: two-layer tier decomposition (alias --ir)
vec4ipa -r <V0,...,V15>     reverse: vectors -> IPA
vec4ipa -n <V0,...,V15>     nearest base segment
vec4ipa -d <A> <B>          weighted distance
vec4ipa -A <IPA1> <IPA2>    sequence (syllable/word) alignment distance
vec4ipa -M FILE             load metric.json weights/lambda at runtime (or --metric FILE)
vec4ipa -D FILE             load custom dimension scheme (ndim/dim/weight/lambda; or --scheme FILE)
vec4ipa -S <CLASS>          reverse symbol class (standard|extipa|sinologist|all; aliases std, ext, school, sino; alias --charset; repeatable)
vec4ipa -P <NAME>           modifier spacing (binary|ternary|2:1:2|1:x:1|X; alias --mode)
vec4ipa -i                  repository / license / feature overview
vec4ipa -R                  full README.md (embedded; or --readme)
vec4ipa -h                  help (usage summary)
vec4ipa -v                  version

I/O (all tools)

• stdin: no input argument — or an explicit - — reads from stdin (echo "tʰa" | ipa2vec, echo "V0,...,V15" | vec2ipa)
• -o FILE, --output FILE: write output to FILE (stderr stays on the terminal)
• -x/-X BASE, --layers-out BASE (alias --ir-out): export the two-layer intermediate representation to BASE.layer1 (character-composition order) and BASE.layer2 (feature-tier order), one token per line: BASE<TAB>ipa<TAB>latin · MOD<TAB>ipa<TAB>latin<TAB>tier · TIE
• -i, --information: repository, license, and feature overview (short CLI summary — the full documentation is -R/--readme)
• -R, --readme: print the embedded full README.md
• --metric FILE: load a metric JSON (the metric.json schema: weights = 16 numbers, lambda, and optionally a full 16×16 metric matrix that overrides weights) and use it for every distance / nearest-neighbour computation in this invocation. Without --metric the compiled-in defaults are used — the binaries never depend on an external file at runtime.
• --: treat every following argument as positional (for strings that start with -)

Examples

$ ipa2vec "tʰeɪk"
[0]    (-0.4500, 0.0000, 0.0000, 0.0000, 1.0000, ...)  pulmonic  [asp
[1]    ...
[2]    ...
[3]    ...

$ ipa2vec -L "ã"           # precomposed char decomposed like a + ◌̃
  layer1 (char order) : [a:front.opn.unr.vwl] → [◌̃:nas/nasal]
  layer2 (feature order): [a:front.opn.unr.vwl] → [◌̃:nas/nasal]
vector[0]: (... vel_open 0.6000 ...)  pulmonic  [nas
rebuilt[0]: /ã/

$ vec2ipa "-0.45,0,0,0,1,0,0,0,0,0.4,0,0,0,0,0,1"
[t]  (vl.alv.pls)  d=0.0000  ->  [t]

$ vec4ipa -q "ʃ"
base: /ʃ/  vl.pst.frc  (pulmonic)
  (-0.3000, 0.0000, 0.0000, 0.2500, 0.6500, ...)

Vectors are in SPEC-NEXT 16-D order — place, body, lips_closed, lips_rounded, tip_shape, tongue_root, vel_open, lateral_ratio, voiced, glottal_aperture, glottal_tension, larynx_height, duration, jet_focus, effective_oral_area, airflow_direction (see docs/SPEC-NEXT.md).

Transcription brackets: /…/ marks the phonemic level (input / alignment display); the reverse fit is phonetic […], and --narrowness narrowest (level 4) prints the narrowest transcription in ⟦…⟧.

Internal logic (two-layer IR)

The parser follows a strict pipeline:

1. Precomposed characters (ã, ẽ, ọ …) are canonically decomposed into base + combining sequence, so ã and a + ◌̃ produce identical vectors.

2. Character composition — the lexer emits tokens in the order they appear in the input string (layer 1): base segments (longest-prefix match against the 133-entry table), modifier letters, ligature ties (͡, ͜, ͠), and preposed modifiers (ᵑǃ, ˀa).

3. Feature ordering — tokens are re-sorted into the natural-language (layer 2) order by feature tier:

   airstream → laryngeal → place → manner → nasal → timing
   

   and applied to the vector in that order (order matters, e.g. asp then creaky).

Every token carries a latin identifier — a scholarly feature name, not the symbol itself and not a folk name:

IPA	latin		IPA	latin
p	vl.bil.pls		t͡ʃ	vl.pst.afc
b	vd.bil.pls		ʘ	vl.bil.clk
s	vl.alv.frc		i	front.cls.unr.vwl
ʃ	vl.pst.frc		a	front.opn.unr.vwl
m	vd.bil.nas		ʰ	asp

Abbreviations: vl voiceless, vd voiced, pls plosive, nas nasal, trl trill, tap, frc fricative, lat lateral, apx approximant, afc affricate, clk click, ejt ejective, imp implosive, bil bilabial, lbd labiodental, den dental, alv alveolar, pst postalveolar, rfl retroflex, alvpal alveolo‑palatal, pal palatal, vel velar, uvu uvular, pha pharyngeal, epl epiglottal, glo glottal, cls close, ncls near‑close, cmid close‑mid, omid open‑mid, nopn near‑open, opn open, unr unrounded, rnd rounded, vwl vowel.

The full name table is data in src/names.tsv — edit it and re-run make gen.

Reverse direction

vec2ipa maps a vector back to IPA: find the nearest base segment (weighted Mahalanobis, including EXTRA_BASE entries), then greedily add the modifier that most reduces the residual distance. Verified: all 133 base segments round-trip losslessly (forward → reverse → forward reproduces the vector to ≤ 0.02 per dimension). Modifier reconstruction is approximate for combinations the greedy search cannot separate (e.g. a ligature that is not a table entry).

Transcription narrowness (--narrowness)

How many diacritics the reverse fit keeps is controlled by --narrowness LEVEL (default narrow; alias --width, which also accepts the legacy levels 0-4): each level sets the maximum number of modifiers per segment and the minimum relative distance gain a modifier must achieve to be kept.

level	name (default)	max mods	min gain	style
0	broadest	2	25%	phonemic, few marks
1	broad	3	10%	broad
2	medium	4	4%	medium
3	narrow	6	1.5%	narrow
4	narrowest	10	0.1%	keep almost every mark

Example (t̬˞̩ˤ → vector → back): --narrowness broadest (or --width 0) gives /ɾˤ˞/, --narrowness narrow (or --width 3) gives /ɾ̝ˤ̆˞/. (Place the option before -r/-n — those consume the vector argument that follows them.)

Modifier spacing (--spacing)

How far an incremental diacritic (raised/lowered ̝̞, advanced/retracted ̟̠, more/less rounded, palatalised, nasalised, …) moves its dimension value is controlled by --spacing NAME (default binary; alias --mode). The step on each side of the neutral 0.5 is 0.5·2/(2+X) for a 1:X:1 ratio:

name	X	step ratio	example (i raised/lowered)
binary	0	1:1	i̞ ≡ e̝ → 0.500 / 0.500
ternary	1	1:1:1	i̞ → 0.533, e̝ → 0.467
2:1:2	0.5	2:1:2	i̞ → 0.520, e̝ → 0.480
1:x:1	x	1:x:1	any x 0–10 (1:2:1, 1:4:1, …; a non-1:x:1 triplet such as 1:2:3 is rejected)
bare X	X	1:X:1	number 0–10, e.g. --spacing 2

Set-type modifiers (nasal, voicing, aperture, …) set fixed values and are not affected.

Supported notations

• All 133 base segments of tools/data/spec_next.scheme (vowels, consonants, affricates, co‑articulated, ejectives, implosives, clicks).
• ExtIPA base segments not in the table: ʬ ʭ ʩ ʪ ʫ ꞎ ᶑ ᴇ ɚ ɞ ɝ (see EXTRA_BASE in src/ipa2vec_core.h). The reverse direction emits standard IPA by default; --symbols extipa adds ʬ ʭ ʩ and --symbols sinologist adds ᴇ ȶ ȡ ȵ ȴ (repeatable, any combination; alias --charset).
• ASCII alias: Latin g == IPA ɡ.
• Diacritics of §10: nasalised ̃, long ː, half‑long ˑ, aspirated ʰ, creaky ̰, breathy ̤ (U+0324), pharyngealised ˤ/̴, velarised ˠ, palatalised ʲ, labialised ʷ, syllabic ̩, non‑syllabic ̯, unreleased ̚, voiceless ̥, voiced ̬, nasal‑click ᵑ (preposed), ejective ʼ (U+02BC, sets glottal_aperture=-1 laryngeal_tension=0.6 voiced=0 and larynx_height=+1; no airstream forcing — the contrast with the base stop lives in the vector, like v8), macron ◌̄ (U+0304, level tone).
• extIPA/clinical marks: dental ̪ (lingual: dentalise; labial: the labiodental stop/nasal p̪ b̪ m̪ / ɱ, encoded on the place axis like t̪ θ ð), linguolabial ̼, laminal ̻, raised ̝/˔, lowered ̞, advanced ̟, retracted ̠, more/less rounded ̹/̜, bridged ͆, apical ̺, ATR ̘, RTR ̙, denasal ̻, mid‑centralised ̽, rhotacised ˞, extra‑short ̆, fortis ͈, lenis ͉, alveolar ͇, whistled ͎, labiodental ᶹ, sliding ͢.
• Ligature ties ͡ ͜ ͠ — table entries (t͡ʃ) match directly; out‑of‑table pairs (tɬ) combine closure + release by the affricate rule.
• Implicit (no-tie) affricates per the IPA affricate table: ts dz tʃ dʒ tɕ dʑ ʈʂ ɖʐ tɬ dɮ cç ɟʝ p̪f each become a single affricate segment. Other letter pairs (gp) stay separate.
• Precomposed Unicode characters (ã ẽ ĩ õ ũ ỹ ạ ẹ ọ ụ é è ê ě ē ā ō ī ū ȅ) decompose to base + combining marks.
• Tone letters ˥˦˧˨˩/꜒꜓꜔꜕꜖ (5‑level), Chinese tone classes ꜀꜁꜂꜃꜄꜅꜆꜇, upstep ꜛ, downstep ꜜ, global rise ↗/fall ↘, and pitch diacritics — see Tone system.
• Superscript letters as modifiers (ʳʴʵʶ ᵉᵌᵤ ᵢᵣ ᴬᴮᴼᴾᵁᵂ ᵊ ⁿ ˡ ˢ), releases (ˀ glottal onset, ʱ breathy/aspirated), offglides (ᶷ ᶣ ʸ), stress marks ˈ ˌ (ignored), linking undertie ‿ (ignored).
• Dotless letters: ȷ→j; ı is ɪ when bare (obsolete, warning) but plain i when followed by diacritics. Rebuilt spellings keep the dotless form whenever a mark covers the dot — the voiceless ring moves above on i/j (i̥→ı̊, j̥→ȷ̊), and any above mark is written on the dotless letter (ĩ→ı̃).
• Invalid combinations parse with a warning (input tolerance); all issues of one segment are merged into a single line, e.g. 4 invalid combinations on i: voiced mark on an already-voiced letter, voiceless and voiced, syllabic on a vowel, syllabic and non-syllabic. Detected classes: syllabic on a vowel (i̩), non-syllabic on a consonant (n̯), the voicing ring on an already voiceless/voiced letter (p̥, b̬), a repeated mark (iːː, n̩̩, ə˞ʳ), a place/secondary-articulation diacritic that repeats the letter's own feature (ʈ̢, ɲ̡, kˠ, ɲʲ, ɚ˞) or contradicts it (t̢̪, ʈ̪, ɲ̪), and contradictory mark pairs on one segment: voicing (i̥̬), timing (iː̆, iːˑ, iˑ̆), phonation (i̤̰), rounding (i̹̜), place shift (i̟̠), height (i̝̞), tongue root (i̘̙), tension (i͈͉), syllabicity (i̩̯), aspiration (iʰ̚, iʰʽ, pʰʱ), place (t̢̪, t̢͇, t̢̡, t̡̼), tip shape (t̺̻) and secondary articulation (lˠˤ). Legitimate derivations stay silent (t̪, t̠, θ̼, ɺ̢, ɲˠ, kˤ).

Unknown symbols produce a U+XXXX error with the byte offset.

Tone system (5-level)

Tone letters after a segment are grouped into a 3-slot annotation printed after the vector as tone=(g1)?(g2)?(g3) with ? as the unknown placeholder and trailing empty groups omitted:

input	meaning	output
ma˩˨	2 letters, single tone	tone=(1,2)
ma˥˧˩	3 letters, 3‑degree single tone	tone=(5,3,1)
ma˩˨꜓꜒	4 letters, single tone + tone sandhi	tone=(1,2)?(4,5)
maꜛ	upstep	tone=?()?(-1,0,0)
maꜜ	downstep	tone=?()?(1,0,0)
ma↗	global rise	tone=?()?(0,1,0)
ma↘	global fall	tone=?()?(0,-1,0)
꜅	Chinese tone class 6	tone=?()?(0,0,-3)

Slot order (authoritative): 1 = single tone / contour (5-level letters ˥˦˧˨˩ or superscript digits), 2 = tone sandhi (꜒꜓꜔꜕꜖), 3 = 3-D vector (upstep, global, class) — upstep ꜛ (negative) / downstep ꜜ (positive), global rise ↗ / fall ↘, Chinese tone class ꜀꜁꜂꜃꜄꜅꜆꜇ mapping to 1, -1, 2, -2, 3, -3, 4, -4 (阴平 阳平 阴上 阳上 阴去 阳去 阴入 阳入). When several tone marks co-occur they parse in this order; the reverse fit prints them as 5-level letters → sandhi → upstep/downstep → global → class. Tone marks bind to the preceding segment.

Regional / tradition modules

Deprecated and regional symbols are organised into modules in src/ipa2vec_core.h (the registry ALIAS_MODULES lists them). School-of-linguistics modules (americanist, sinologist, indologist, polish, teuthonista, koreanologist, japanologist, africanist, oed) resolve their symbols by default but print a warning listing every school that contains the symbol:

ipa2vec: warning: using symbol 'ł' from americanist, polish — enable with --americanist --polish

Pass --<name> (e.g. --polish) to use that school's readings without a warning. When a symbol exists in several enabled schools, the command-line order decides: --polish --americanist ł gives Polish [w], --americanist --polish ł gives Americanist [ɬ]. generic, withdrawn, equiv and uppercase are always on and never warn.

module	symbols
generic	?→ʔ, я→ʢ, ∅ Ø null, ȷ ɫ ı dotless/dark, small-capital glyphs
withdrawn	ƍ σ ƺ ƪ ƻ ƾ ʦ ʣ ʧ ʤ ʨ ʥ ʇ ʗ ʖ ʞ ƥ ƭ ƈ ƙ ʠ ƞ ƫ ʓ ʆ ɼ ɩ ʚ ɷ ω ȣ
americanist	š č ž ǰ ǧ ǯ ẋ ƛ ł λ
sinologist	ɿ ʅ ʮ ʯ ᴀ, plus base segments ȶ ȡ ȵ ȴ (alveolo-palatal stop/nasal/lateral, gated on --sinologist)
indologist	ḍ ṭ ṇ ṛ ḷ ṣ ś ṃ ṅ ñ ḥ ḫ ẓ ẖ ḏ ṯ ġ ḡ ḻ ṟ ṁ
polish	ź ć ż, plus Polish values for č š ž ł (conflict with Americanist)
teuthonista	ƀ đ ǥ ǩ ȟ ǵ
koreanologist	Ǝ→ɤ, K P T fortis
japanologist	Q sokuon, N syllabic nasal
africanist	ȹ ȸ
oed	ᵻ ᵿ
uppercase	G→ɢ, R→ʀ, Œ→ɶ (only explicitly listed; no shape extrapolation)

Inference reporting

Every inference the parser makes is reported to stderr:

• note: inferred tie: 'ts' -> 't͡s' — implicit (no-tie) affricates merged
• note: inferred affricate with synthesized tie — out-of-table pairs (tɬ, dɮ, cç, ɟʝ, p̪f)
• note: ASCII 'g' interpreted as IPA ɡ
• note: ... reinterpreted ... — symbol reinterpretation: ASCII ' → unreleased, ′ → palatalization (Irish), ʻ ‘ ‛ → weak aspiration, ʱ → breathy/aspirated release
• note: '<alias>' -> <repl> (<name>) — any deprecated / non-standard mapping
• warning: '<alias>' -> <repl> (<name>) — deprecated or ambiguous usage where the formal reading was preferred (ʞ→ǃ, Q→gemination, K P T→fortis, bare ı→ɪ)

Stress tests

Run the full suite:

python3 tools/test_suite.py       # 217 checks: parsing, tone, regional,
                                  # inference, warnings, round-trip

Additional suites (all against the built binaries unless noted):

python3 tools/test_metric_space.py      # 681 checks: nearest-base anchors,
                                        # metric-space round-trip (needs ipa2vec+vec2ipa)
python3 tools/test_alignment.py         # 57 checks: curve-distance alignment
                                        # (absorption, vowel blocks, tone; needs ipa2vec)
python3 tools/validate_alignment.py     # axiom fuzz: symmetry, identity,
                                        # absorption licensing (needs vec4ipa)
python3 tools/test_standard_chinese.py  # 448 checks: Mandarin initials/finals/tone
python3 tools/test_spec_next.py         # 121 checks: 16-dim table + metric JSON
                                        # (no binaries needed)
python3 tools/verify_modifiers.py       # modifier model sanity (design model,
                                        # no binaries needed)
python3 tools/fuzz_metric_space.py      # random-vector fuzz over the metric

The clinical/extIPA stress string — every feature in one input (glottal onset, rhotacised vowel, apical, undertie, stress, mid-centralised, offglides, pharyngealisation, dentalisation, fricative release, syllable break, downstep, nasalised linguolabial, half-long, tone classes, 5-level tones, tone sandhi, global rise, prosodic breaks, creak, breathy, dotless i, diaeresis, labio-palatal offglide, syllabic, lateral release, non-syllabic, nasal release, denasal, palatalised, retracted, more-open, rhotacised, lowered, velarised, unreleased, extra-high tone, more-close, voiced, voiceless ring, retroflex hook, precomposed accented vowels, labiodental plosives (ȹ ȸ), alveolo-palatal (ȶ ȡ), ejectives, clicks, implosives, lateral affricates, pitch contours):

ˀɝ̺̆k‿ˈo̽ːᶷˌre̴̪pstˢ.ꜜã̼ˑd꜅˩˦˧꜕꜖꜒↗|ba̜˔z̟ʱ.tsᵊn↗‖z̥ı̤̃ˤt̬ṵ̈ːdᶣ.r̥ʰl̩pˡ.hr̯ʱdⁿ|
βʷᵿ̻ʑ̩ʲːn̠↘‖ø̙˞˕dˠ̚‖ʎ̯e̘̋t̬˞̩ˤŋ̊.ɺ̢ é̙dɮ.ȹē̹ʟʈʼ.ȸè̜.ȶʎ̝̥ʼȅ̞ȡ.kʟ̝̊ʼě̝ʡ̯.tɬʼḛ̂ʙ.
ʘe̤᷄ɗ.χʼe̥᷅ǃ.ʛe᷇ǂ.ɧe᷆t͡s.ʑe᷉ɡ͡b.ŋ͡mʍ̩ɥe᷈ɫ

ipa2vec "<above>" parses all 76 segments; the tone annotation on the d꜅˩˦˧꜕꜖꜒↗ group reads tone=()?(1,4,3)?(2,1,5)?(?,1)?(-3).

Files

File	Purpose
src/ipa2vec_core.h	shared core: UTF-8, modifiers, aliases (modules), lexer, canonicaliser, applier, reverse fit
src/ipa2vec_main.c	ipa2vec — IPA → vectors (parse, IR, JSON)
src/vec2ipa_main.c	vec2ipa — vectors → IPA (nearest, reverse fit, distance)
src/vec4ipa_main.c	vec4ipa — inventory + both directions
src/vectors.h	generated: 133 base segments + metric weights + latin names
src/names.tsv	data: symbol → scholarly feature name (edit + make gen)
src/readme_embed.h	generated: this README embedded for vec4ipa -h
tools/gen_vectors_h.py	regenerates src/vectors.h
tools/gen_readme_embed.py	regenerates src/readme_embed.h
tools/test_suite.py	217-check regression suite (incl. the stress string)
tools/test_alignment.py	57-check curve-distance alignment suite
tools/validate_alignment.py	property-based axiom fuzz (symmetry, identity, absorption licensing)
docs/SPEC-NEXT.md	the current 16-D vector specification
docs/ALIGNMENT.md	the -A sequence-alignment design: absorption gates, vowel blocks, the unified mora axiom
docs/SPEC.md	the legacy v8 16-D vector specification (historical)
IPA_VECTORS.md	legacy v8 vector table (historical; the current table is tools/data/spec_next.scheme → src/vectors.h)
tools/data/spec_next.scheme	the 133-segment vector table (data for the generator)
METRIC.md	metric derivation (weights, λ)
metric.json	machine-readable weights + λ (v9)
Makefile	build all three tools (auto -municode on Windows)

License

MIT — see LICENSE.