In September 2024, the work of the Sustainable Innovation and Technology Foundation specialists was focused on several aspects of the KazLLM project development.

First, the development of a new tokenizer was completed, it provides more efficient training on Kazakh text. The KazLLM model was retrained, datasets in Kazakh, Russian and Turkish were added to it, and this will allow the model to work with a wider language audience and improve the quality of responses.

Second, a vocoder for speech translation was configured and trained, and it will allow translation from Kazakh in real time. The Soyle product was launched, which is already available at soyle.nu.edu.kz. It implements text-to-speech and text-to-text translation functions, and can process up to 2000 characters in one request.

Third, as part of the work on the development of digital avatars, all components of the system were combined into a single solution on one computer. The system is now capable of autonomously generating responses and displaying them via a virtual avatar, with response time reduced to 1.46 seconds. This integration has significantly accelerated the system and improved user experience.

The implementation of work within the framework of the KazLLM development project was made possible with partial support from AstanaHub.