The article presents work on predicting the fundamental frequency (F0) values for the Kazakh language. The fundamental frequency plays one of the most important roles in the perception of speech, and at the same time modelling continuous F0 is one of the most difficult tasks in the development of intonational speech synthesis systems. The main and obvious difficulty is that a person is able to say the same sentence with different intonations and with different tones. In this work, we used deep neural networks for accurate and qualitative prediction F0 values as close as possible to the natural sounding of Kazakh speech.