<p>Attention mechanisms have demonstrated substantial effectiveness in recent deep learning-based speech enhancement models. Transformer-based techniques surpass conventional neural networks, such as RNNs and CNNs, in managing long-term dependencies. To address the limitations of prior models, a frequency transformer is incorporated that performs spectral convolution within the F-Transformer module. The proposed speech enhancement architecture, SCTCFT-STCM-TAN Net, integrates sub-convolutional T-Conformer F-Transformers, squeezed temporal convolutional network modules, and transformer attention networks. Employing a sub-convolutional encoder–decoder algorithm with variable kernel sizes facilitates the extraction of multi-scale local and contextual information from noisy speech. The architecture leverages a U-Net encoder and decoder for sequence modelling and enhanced feature learning, utilizing a transformer-based attention network that links the sub-convolutional T-Conformer with compressed temporal convolution and F-transformer spectral convolution. The model exploits the time–frequency structure of spectral components to enhance output quality and incorporates a dynamic hierarchical attention module, along with multiple adaptive time–frequency attention components, to capture stacked contextual information and long-term dependencies. STCM blocks are dedicated to temporal sequence modelling. Experimental results indicate that the proposed model outperforms several state-of-the-art approaches. The evaluation of the proposed model is conducted using short-time objective intelligibility (STOI), signal to distortion ratio and perceptual assessment of speech quality (PESQ) metrics. The model demonstrates superior performance compared to existing techniques. Average PESQ and STOI scores on the Common Speech dataset are compared against those of noisy speech models across various noisy environments.</p>