Proto-Indo-European is the ancestor of most languages spoken in Europe, Iran and northern India today. It was spoken around 4000 BCE but never written down, so everything we know about it has been pieced together by comparing its descendants. Each reconstructed form is marked with an asterisk (*). This project builds the language as a working system: a lexicon; a grammar engine that inflects words by the reconstructed rules; texts drawn from scholars' reconstructions or composed through the engine and checked against those same rules (each record says which); a language model fine-tuned on this material, so it can read and write Proto-Indo-European itself; and synthetic voices trained to say it aloud.
The homeland was probably north of the Black Sea and the Caspian, in what is now Ukraine and southern Russia: the Pontic–Caspian steppe. Its people herded cattle, sheep and goats, and drove wagons on solid wheels (not spoked ones) pulled by oxen. They kept horses, and their language had a word for them, *h₁éḱwos. Whether they rode them is still argued over.
Over the next two thousand years or so their descendants moved west into Europe, south into the Balkans, Greece and Anatolia (modern Turkey), and east through Central Asia to India, where one group became the Vedic people who composed the Rigveda in Sanskrit. As they went, their speech split into branches, and each branch became a separate language.
The earliest texts in any of these languages are Hittite: clay tablets in cuneiform from Hattusa, the Hittite capital, from the 1600s BCE. The date beside each branch on the map is its earliest text.
No branch ever wrote down Proto-Indo-European itself; by the time we have texts at all, it had been gone for two thousand years.
But because languages change in regular ways (the same sound change happens to many words at once), scholars can work backwards. Take the word for father: Vedic Sanskrit pitár, Ancient Greek patḗr, Latin pater, Gothic fadar (in the Bible translated by Wulfila around 350 CE). These all come from a single form, *ph₂tḗr. The asterisk means reconstructed by comparison, not directly attested. Every word in this project's lexicon is marked the same way.
What reconstruction recovers least well is syntax: word order, how questions were formed, the particles that glued clauses together. Some of it is fairly secure, like *-kʷe “and” tacked onto the end of a word, or the verb tending to come last. Much of the rest is inferred from the oldest daughter languages and remains uncertain.
The corpus is not recovered text. It is PIE composed to follow the reconstructed grammar. Each record states its source: a published scholarly reconstruction, an attested poetic formula, a back-projection from parallel phrases in the daughter languages, or a passage generated and checked by the engine.
We tested our model against frontier models on a fixed set of questions held out of all training: inflection, full paradigms, translation in both directions, and dialogue. Each answer gets two scores. The lenient score ignores notation, so a form without accents or with ASCII laryngeals still counts. The strict score requires the standard scholarly notation. The frontier models score well on the lenient measure and lose 25 to 36 points on the strict one. Our model loses about two.
PIE score, 0–100, weighted over five segments. Frontier models answered without tools; each is shown at its best effort level. Ours is Nemotron 3.5 Super (122B), fine-tuned, stage e3.
This hymn to the dawn goddess *h₂éwsōs comes from the corpus. Her name survives as Vedic Uṣás, Greek Ēṓs and Latin Aurōra. It is read by the project's own PIE voices, Kokoro text-to-speech models we trained for the language. Their training speech began in a simulated vocal tract (VocalTractLab, an articulatory synthesizer), where each sound is a movement of tongue, lips and larynx: the uvular fricative of h₂, the palatal stops, the breathy-voiced stops, the pitch accent. Speech-to-speech conversion then carried it into natural human voices.
The translator below runs on the project's grammar engine. It parses English, picks the PIE words, inflects them, and gives a gloss and pronunciation for each. A chat interface to the model and a reader for the corpus are in progress.