Kryptonit, a subsidiary of ICS Holding, has developed a new fine-tuning methodology for speech-to-text models. This approach improves the accuracy of transcribing mixed Russian and English speech within a single audio segment. During internal testing, the company recorded a 13.6% reduction in the word error rate compared to the baseline open-source model.
To achieve these results, the team prepared a dataset consisting of approximately 10,000 hours of audio. The data was generated by combining existing language segments and synthesizing speech where sentences switch between languages. This training allows the model to correctly identify and transcribe English terms in Latin script rather than transliterating them into Russian.
According to Artem Ryzhenkov, head of audio processing at Kryptonit, this method demonstrates that existing open-source models can be adapted for complex scenarios without the need to build new architectures from scratch. The developers noted that while the current experiment focused on Russian and English, the methodology is adaptable to other language combinations.
The global market for speech recognition APIs is projected to reach $7.21 billion by 2031, with a compound annual growth rate of 20.23%. Industry demand is driven by the increasing use of automated transcription in corporate meetings and customer service environments, where language switching is common.
