Does AI speak Erzya? Translation has arrived, the voice has not

A year ago, in July 2025, Erzya and Moksha — the two Finno-Ugric languages of the Republic of Mordovia in central Russia — were added to Yandex Translate, Russia's leading translation service, becoming its eleventh and twelfth languages of the country's peoples. The parallel corpus behind the launch, more than 220,000 sentence pairs, was prepared by the philology faculty of Ogarev Mordovia State University in Saransk. The anniversary is a good moment to survey the whole landscape: where neural networks already understand Erzya, and where the language simply does not exist for them.

The road to Yandex was longer than it looks. The computational infrastructure of Erzya has been under construction since the 2000s in Tromsø and Helsinki: the Giellatekno project and the Finnish linguist Jack Rueter built the morphological analyser, electronic dictionaries and corpora on which almost everything later rests. The first neural translator for Erzya was created in 2022 by the researcher David Dale together with the Erzya language community — native speakers judged more than half of its output acceptable. In 2023 Erzya and Moksha entered Neurotõlge, the University of Tartu's translation engine covering 23 Finno-Ugric languages, and in 2024 the same team of enthusiasts — Isai Gordeev, Sergey Kuldin and David Dale — translated the international benchmark FLORES+ into Erzya and presented the results at the WMT24 conference: for the first time, Erzya was plugged into the world's standard system for evaluating machine translation.

Against this background, the gaps stand out. Google Translate, which added 110 languages in one sweep in June 2024 — including Komi, Udmurt, Mari, Chuvash and Bashkir — passed over Erzya and Moksha; nor do they appear in Meta's NLLB-200, the model that promised to "leave no language behind". Two languages with hundreds of thousands of speakers have so far found no place in the global services; open models fine-tuned by the community stand in for them.

With speech, the divide is sharper still. The only publicly available systems for synthesising and recognising Erzya speech are Meta's open MMS models from 2023, trained largely on recordings of the Bible: a single voice with an ecclesiastical intonation. Commercial services do not hear Erzya at all: it is absent from OpenAI's Whisper recogniser and from the voice assistants Alisa, Marusya and Siri, and even Yandex, having added translation, did not switch on speech features for Erzya — as of July 2026, its voice input and read-aloud work for Tatar, Chuvash, Bashkir, Mari, Udmurt and Yakut. Neural networks can now read Erzya text; Erzya speech, in effect, they still cannot.

The universal chatbots have learned Erzya by halves. The WMT24 evaluations showed that large language models understand Erzya better than they write it: Claude translated Erzya text into Russian more accurately than any other system tested, but in the opposite direction it lost to the specialised model fine-tuned by the enthusiasts. A study presented in December 2025 at IWCLUL, the workshop on computational Uralic linguistics, added a telling detail: GPT models refused outright to translate into Erzya in 27% of cases. You can already ask a neural network about Erzya; getting a well-formed Erzya answer from one is another matter.

A simple conclusion follows. A language exists for artificial intelligence exactly to the extent that corpora have been collected for it. The data behind Yandex was created by a university, the benchmark by three researchers and a community, and the speech models rest on a Bible translation. Every digitised text and every hour of recorded speech — for instance in the open Common Voice project, which accepted Erzya back in 2022 — widens the territory on which the language is visible to machines. And the reverse holds too: what is not in the data does not exist for the neural networks.

Valks sees this as a direct continuation of its own work: an open dictionary, phraseology, texts and audio materials are precisely the data from which tomorrow's Erzya translators, speech synthesisers and voice assistants will grow. We will keep following how neural networks take up Erzya — and we would add a reminder: the only people who can teach a machine to speak Erzya are those who speak it themselves.