In this paper the comparison of several types of signal transform layers in speech enhancement using deep neural networks was done. The main requirements to be considered when choosing the type of transform were the quality of the enhanced signal, the size of the neural network model and the possibility of real-time operation. It was proposed to apply recurrent transformations of LSTM type and two types of state space transformations S4 and Mamba. The neural network consisted of two main blocks of transformations. The first one processed the whole spectrum of the signal and the second one worked with separate frequency bands. Data from the DNS Challenge set were used as training data. They were used to generate examples of distorted signals with signal-to-noise ratio in the range of 0–20 dB. The trained models were tested on a test subsample of the same set, as well as on synthesized samples with a given value of the signal-to-noise ratio in the range from −5 to 20 dB, in order to assess the generalization ability of the models. The results of numerical experiments showed that the LSTM-based and Mamba-based models performed the best. The latter had a significantly smaller number of trained parameters, which allowed us to consider it as the most suitable for improving the quality of speech signals.