Objectives. The development of contemporary models for the conversion of accents in foreign languages utilizes deep neural network architectures, as well as ensembles of neural networks for speech recognition and generation. However, restricted access to implementations of such models limits their application, study, and further development. Moreover, the use of these models is limited by their architectural features, which prevents flexible changes from being carried out in the timbre of the generated speech and requires the accumulation of context, leading to increased delays in generation, making these systems unsuitable for use in real-time multiuser communication scenarios. Therefore, the relevant task and aim of this work is the development of a method that generates native-sounding speech based on input accented speech material with minimal delays and the capability to preserve, clone, and modify the timbre of the speaker’s voice.Methods. Methods for modifying, training, and combining deep neural networks into a single end-to-end architecture for direct speech-to-speech conversion are applied. For training, original and modified open-source datasets were used.Results. The work resulted in the development of a real-time accent conversion method with voice cloning based on a non-autoregressive neural network. The model comprises modules for accent and gender detection, speaker identification, speech conversion, spectrogram generation, and decoding the resulting spectrogram into an audio signal. As well as demonstrating high accent conversion quality while maintaining the original timbre, the short generation times of the applied method make it acceptable for use in real-time scenarios.Conclusions. Testing of the developed method confirmed the effectiveness of the proposed non-autoregressive neural network architecture. The developed model demonstrated the ability to work in real-time information systems in English.