The Doctrine of Sound
AI voices flatten the gods into English reading rules. This is the mechanical system that says each name the way it was meant to be said — the same marks we restore, decoded into sound and time.
The Meta-Rule
Every restored name on PuniCodex carries its own speaking instructions. Three marks do all the work — and they are read mechanically, never by editorial mood.
Stress this syllable. Zeús is ZE-oos, never "zoose". The accent is where the voice rises.
Hold the vowel double — one extra beat of time. Apóllōn holds the final ō twice as long as the rest.
Stress and length in one mark, with a falling pitch contour across the long syllable — Athēnâ's final vowel.
The Timing Model
A name is not a string of sounds but a string of time. Each syllable earns morae — the unit classical Greek, Japanese, and Sanskrit all independently use:
| Shape | Morae | Example |
|---|---|---|
| Short vowel | 1 | a in Apóllōn — a flicker of time |
| Long vowel / diphthong | 2 | ō in Apóllōn — double the hold |
| Geminate consonant | +1 | ll in Apóllōn — the held liquid |
| Heavy coda | +0.5 | closed syllable tails, rounded per syllable |
Read our beat strings as: ˘ = one mora, ¯ = two, ¯˘ = three. Apóllōn = ˘ ¯ ¯˘ — 1, 2, 3 morae, ~660 ms, stress rising on the middle syllable.
The Language Tables
The same tokenizer reads the restored form and applies the language's own rules — no English assumptions anywhere.
| ā ē ī ō ū | long vowels — ē = open "eh" held, ō = "aw" held |
| acute / circumflex | stress (circumflex adds length + falling pitch) |
| z · x · ps | zd · ks · ps — Zeús begins with the voiced affricate |
| ph th kh | aspirated p/t/k — a breath after the stop, never "f/th/x" |
| final -ē, -e | always sounded — "Hekátē" ends in "tay", never silent "e" |
| tl | one affricate — the "whispered l" /t͡ɬ/, never "tul" |
| tz | /t͡s/ as in "cats" |
| qu (e/i) | plain /k/ — Quetzalcōātl starts "keh", not "kwet" |
| hu | /w/ before vowels |
| stress | penultimate syllable, unless marked otherwise |
| þ | /θ/ — "th" of think: Thór |
| ð | /ð/ — "th" of breathe: Óðinn |
| ǫ | /ɔ/ — the open o of Ragnarǫk |
| acute vowels | long: á é í ó ú ý |
| stress | first syllable, Germanic initial |
| ś ṣ | ɕ / ʂ — two distinct sh-sounds: Śiva vs Viṣṇu |
| ṭ ḍ ṇ | retroflex — tongue curled back |
| ṛ | syllabic r — a vowel made of r |
| bh dh gh th ph | breathy stops — h is a feature, not a letter |
| ā ī ū | long vowels, counted as mātrā |
| ō ū | double-beat vowels — Kyōto is Kyo-o-to, four beats |
| r | ɾ — a single tap, between l and d |
| sh ch ts | ɕ · t͡ɕ · t͡s |
| rhythm | strict isochrony — every mora equal (~80 ms) |
Hieroglyphs write consonants, not vowels. Egyptological readings (Rꜥ as "Ra", Ḥr as "Hor") are conventions of the field, not derivations — and they are labelled "Conventional reading" wherever they appear. We do not fake certainty the script doesn't offer.
For Voices
Every flagship temple carries a "Say it right" panel: the respelling (stressed syllable in CAPS), the IPA, the beats, and the one or two sounds English mouths get wrong.
AH-rays · /ˈa.rɛːs/ · ˘ ¯˘ · risesharp first syllable, long open "eh" at the endkeh-tsahl-KOH-ahtl · ˘ ¯ ¯ ¯˘ · heavytl is one sound; the ō and ā each hold doubleOH-theen · ¯ ¯ · heavysoft "th" of breathe, never a hard dah-mah-teh-rah-soo · ˘ ˘ ˘ ˘ ˘ · flatfive even beats, tapped r, no stress at allFor Machines
Every name in the lexicon ships a machine-ready pronunciation record:
GET /api/v1/names/{id}/pronunciation — IPA, respelling, syllables, stress, SSML markup, and the full timing object (morae, beats, contour, durations).
Each record carries <phoneme alphabet="ipa"> payloads that ElevenLabs, Google TTS, Amazon Polly, and Azure accept natively.
data/corpus/pronunciation.jsonl — 900+ records for training and evaluation, refreshed with every dataset generation.
The Honest Boundary
Most flagship names are derived mechanically from their restored marks (and validated against the curated pronunciation atlas — currently ~82% exact-or-near across 893 entries, 89% mora agreement on flagships). Egyptian readings are labelled conventional. Everything else falls back to the curated atlas rather than a guessed rule. Where the rules end, we say so.