Neural Networks Still Stumble on Belarusian Speech. Belarusians Want to Give Artificial Intelligence the Ideal Voice

Recording studio. Illustrative photo. Photo: Freepik / DC Studio

An important technological breakthrough in Belarusian speech synthesis occurred back in the spring of 2025, thanks to Google's implementation of its new Gemini model, which learned to accurately recognize Belarusian speech (STT — Speech-to-Text), thanks to which, for example, automatic Belarusian-language subtitles finally appeared on YouTube.

Belarusians themselves contributed greatly to this through the volunteer project Donar.by, collecting thousands of hours of live voices.

Thanks to this gigantic database, today it is Google's voice that is closest to the correct sound of the Belarusian language. The model understands context well and has a huge vocabulary, leaving competitors from OpenAI or ElevenLabs far behind, whose attempts to speak Belarusian are far from natural speech.

But recognizing speech is only half the battle. When a neural network has to voice text itself (TTS — Text-to-Speech), it systematically makes mistakes in rarely used words and cannot cope with homographs — words that are spelled identically but have different meanings depending on the stress.

If, instead of the correct "sparyshámi", artificial intelligence confidently produces "spáryshami", this immediately reveals its synthetic nature to a native speaker. The native speaker may not even know the meaning of the word, or where the stress is placed, but their linguistic intuition tells them that something is wrong.

Furthermore, such errors, even if rare in Google models, do a disservice to those who are just beginning to learn the Belarusian language, reinforcing distorted pronunciation.

Add to this the problems with conveying the softness of consonants, the specific sound of "ў", and the affricates "дз" and "дж" — listening to and perceiving long texts in such a rendition is still physically difficult.

Voice from a Test Tube

The problem is not that the algorithms are not smart enough — in the case of the Belarusian language, they simply have nothing to learn from. For artificial intelligence to master correct intonation, rhythm, and stresses, audio from YouTube or podcasts, where sound quality is always varied and people's diction is imperfect, is not enough.

Павялічыць

To create a natural synthesized voice, a special, crystal-clear studio dataset is required. This means thousands of hours of professional readings, where texts are specially constructed by linguists to cover all possible phonetic combinations and show the model how to correctly place stresses in complex contexts. Today, no such open data array exists for the Belarusian language in the world.

It is precisely this empty niche that the Sonora project intends to fill. This is a volunteer initiative driven by project manager Hanna Maklakova, linguistic engineer Uladzislau, the TuteishyGPT development team, and a number of specialists whose names are not disclosed for security reasons. Their goal is not to create a closed commercial product, but to establish a fundamental database that everyone can use.

How They Plan to Create the Ideal Voice

Currently, the team is at the fundraising stage, planning to collect 13,000 euros for the project. The largest part of the budget will go towards renting a professional studio and paying speakers with ideal pronunciation. The rest will cover sound engineers' services, the painstaking work of linguists who will prepare and annotate the text corpus, and other expenses.

The result of this work will be a completely open dataset with a public license. Based on it, the project authors plan to refine the already existing domestic BexTTS model, bringing it to a fundamentally new level.

The team is seeking direct contact with representatives from Google, OpenAI, Meta, and Speechify to offer them ready and high-quality material. In the logic of global corporations, everything is simple: if they are given a ready tool to improve a product in a local market, they gladly integrate it.

If they cannot collect the entire amount immediately, the project authors promise to start recording with the funds already in their accounts, because even a partial replenishment of the database is a practical step forward.

From Textbooks to Navigators

The presence of the Belarusian language in technologies today is a matter of its survival in principle. High-quality speech synthesis fundamentally changes the rules of the game in content creation.

This means that publishing Belarusian audiobooks or voicing long articles will no longer require huge budgets and weeks of studio work. It's an opportunity for schoolchildren and students to listen to textbooks, and for people with visual impairments or dyslexia — to get full access to Belarusian-language information. It's a basis for creating domestic voice assistants, chatbots, and navigators that will not speak to us in broken Google speech. Finally, it is a convenient tool for the enormous Belarusian diaspora that wants to preserve the linguistic environment for their children abroad.

• 20
• 0
• 0
• 1
• 0
• 0
• 0
• 0