In this paper, we considered the possibility of Mongolian speech recognition using a neural network. At the same time, we applied a method with highlighting the syllable of a recognizable word, since in the Mongolian language words are formed by adding a suffix to its root. And the method of extracting the syllable of a word allows you to reduce the size of the database, since there is no need to save the features of each word. The neural network was trained using the Levenberg–Marquardt algorithm, which is more stable and faster when working with a small database. The experiment used 4 words of the Mongolian language with one or two syllables. The words were written directly into the computer using the Matlab software. The speakers who voiced these words are 11 people, men and women of different ages. Of these, 1 person aged 20–30 years, 7 people aged 30–40 years and 3 people aged 60–70 years. In addition, the neural network recognizes the area in speech where the noise occurs in order to distinguish and separate the speech area from the noise. The trained network recognizes the words of the Mongolian language by syllables from the database with a probability of 96.5 %.