The article focuses on the problem of audio signal quality during video conferences. The effect of noise on the quality and intelligibility of the speech signal is described. Analysis of the noise reduction in the audio signal in real time has been carried out. The main problems arising in the digital processing of the audio signal in real time are highlighted. General methods of reducing the noise are considered and the disadvantages of classical methods are revealed. The problem of eliminating noise with a variable frequency band by using classical noise reduction methods is formulated. There is substantiated the need for creating a hybrid noise reduction technique by using machine and deep learning methods to eliminate both static noise and noise with complex and variable spectral characteristics. The main approaches to solving the problem of noise reduction in real time are highlighted, namely the approach with recognition and elimination of noise and the approach with voice recognition and elimination of sounds that differ from the speech signal. A noise reduction algorithm based on an approach with recognition and elimination of noise is described. Optimization of the algorithm is proposed by decomposing the spectrum of the input signal according to the Bark scale. A recurrent neural network is proposed as a tool for implementing a noise reduction algorithm. The formats of the input and output data of the neural network as well as the format of the training data are defined. A model for adjusting parameters and rules for adapting the noise reduction algorithm to the specific operating conditions is described. A hybrid noise reduction technique combining classical noise reduction methods and methods based on a recurrent neural network is proposed. A scheme of a hybrid technique has been developed. A method of testing the effectiveness of the noise reduction technique is proposed.