SPEECH RECOGNITION BASED ON CONVENTIONALNEURAL NETWORKS

In this research work, the problem of speech recognitionis considered in the form of an analysis of the numbers from 1 to 10 recorded by the speakeron the dictaphone. The paper uses the method of recognizing thespectrogram of an audiosignal using convolutional neural networks. Also written and implemented an algorithm for processinginput data, and an algorithm for recognizing spoken words. In this work, the qualityof recognition was assessed for a different number of convolutional layers. A comparison of the recognition quality is made in cases when the input data for the network are the spectrogramof the audio signal or the first two formants extracted from it. The recognition algorithm was tested using examples of male and female voices with different pronunciation lengths.