Belarusian digital voice will no longer grate on the ear with incorrect stresses. The neural network has finally been taught to distinguish "mUka" from "mukA"

As the creators of the open project BelVoice report in their Telegram channel,

the main obstacle to high-quality text-to-speech (TTS) conversion remained homographs. These are words that have the same spelling but different stress: for example, "mUka" (torment) and "mukA" (flour), "vUchyts" (to teach - imperfective) and "vuchYts" (to teach - perfective), "gAdy" (years) and "gadY" (reptiles).

Traditional algorithms used the most frequent variant from the dictionary without any consideration of context. Such an approach provided an accuracy of about 92%. At first glance, the number seems high, but for speech synthesis, this is a critical flaw: an eight-percent error rate means that

practically in every paragraph, the system made a phonetic error that immediately revealed the artificiality of the voice and grated on the user's ear.

To solve this semantic task, researchers integrated large language models (LLMs) into the process. The optimal solution was a digital analysis architecture with role distribution. First, a powerful neural network (e.g., Gemini Pro 3.1) processes data from the Grammatical Database and the Explanatory Dictionary of the Belarusian Language. Based on this array, an optimized prompt is created with clear rules and logical markers for a smaller inference model, which then directly works with the user's text.

All technical instructions for building this algorithm have been openly published by the developers on the GitHub platform.

The selection of the final model for the project was conducted on three specialized datasets: the CommonVoice base, a synthetic balanced 10/10 set, and the first fifteen chapters of Arkadz Charnyshevich's book "Zastsienak Malinauka" with entirely manual stress marking. Detailed research results are published on the Hugging Face platform.

The best ratio of speed and quality was demonstrated by the Gemma 4-31B model without enabling the reasoning function, which excessively slows down performance. On a sample of 1090 complex homographs in natural literary text, the system made only three errors, achieving an accuracy of 99.72%.

Reasoning function — an AI operating mode where it doesn't give an answer instantly but first conducts an internal step-by-step logical analysis. This slightly increases accuracy but catastrophically slows down the work speed (up to a minute per query), so for fast text voicing, this mode is disabled.

For comparison, the competing model Qwen3.8-27B made 25 errors on the same data volume, and Mistral-small-4 made 182 errors with an resulting accuracy of about 83 percent.

• 7
• 0
• 0
• 0
• 0
• 0
• 0
• 0