Deep learning for speech enhancement: architectures, paradigms, and emerging trends

Ivanko D.V. Deep learning for speech enhancement: architectures, paradigms, and emerging trends. Scientific and Technical Journal of Information Technologies, Mechanics and Optics, 2026, vol. 26, no. 4, pp. 673–682 (in Russian). doi: 10.17586/2226-1494-2026-26-4-673-682

A systematic review of modern neural network methods for Speech Enhancement is presented, aimed at improving speech intelligibility and quality under acoustic distortions. The reviewed methods can be applied in voice control systems, telecommunications, hearing aids, and human–machine interaction interfaces. Key architectural approaches are considered, including classical recurrent and convolutional networks as well as modern hybrid architectures with attention mechanisms (Transformer, Conformer), state-space models (Mamba), and advanced recurrent blocks (xLSTM). The advantages and disadvantages of different architectures are shown in terms of speech restoration quality and computational efficiency. Specific problems of existing methods are highlighted, including high computational cost and insufficient generalization capability under non-stationary noise conditions. The need for further research in the development of lightweight models for mobile devices, multi-distortion suppression methods, and the integration of neural network noise suppression with generative models to achieve a new level of speech signal restoration quality is demonstrated.

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License