The model is capable of accurately recognizing languages such as Kazakh, Uzbek, Kyrgyz, Turkish, Azerbaijani, and Tatar, as well as processing mixed speech that includes elements of both Turkic and Russian languages. This will improve the quality of voice services for government and commercial organizations.
Previously, international ASR models focused on the English language were used in Central Asia, which reduced recognition accuracy due to the peculiarities of Turkic phonetics and pronunciation. The new development, created from scratch for natural conversational speech, demonstrates high resilience to background noise and accents, and effectively handles switching between languages within a single conversation.
The project was implemented using Microsoft’s GPU infrastructure and received strategic support from Astana Hub and the Ministry of Digital Development, Innovations, and Aerospace Industry of Kazakhstan.
