Russian Researchers Cut Speech Recognition Errors in Half
Engineers at Russia’s Sber have unveiled a new method that slashes AI speech recognition mistakes by 50 percent — a leap that could reshape how machines understand human language.
At the Interspeech 2025 conference, Sber’s team introduced a model that rethinks how artificial intelligence interprets spoken Russian. The method draws on target variables from a Connectionist Temporal Classification (CTC) framework, a shift that makes the system generate more precise semantic representations of audio data.
In testing, the approach cut error rates in half compared with earlier solutions.
“This method shifts the focus to semantic representations and shows high efficiency,” said Fedor Minkin, CTO of Sber’s GigaChat project
The method could be particularly useful for developing automatic speech recognition systems for languages with limited training data. According to Minkin, “the development has the potential to become a new standard for the industry and lay the foundation for the next generation of voice interfaces.”
For now, the Sber team’s work stands out as a major step forward in global speech-recognition research.
