GT-LFP: a progressive speech enhancement network with selective state space and dynamic frequency awareness

Speech enhancement is a core issue in speech signal processing. It holds significant value in speech communication and speech recognition. This paper addresses the challenge of balancing model complexity and performance in deep learning-based speech enhancement. A Progressive Speech Enhancement Network, based on selective state space and dynamic frequency awareness, is proposed. The model uses a Selective State Space Enhancement (LSSE) module, a Frequency-Aware Dynamic Convolution (FADC) module, and a Progressive Mask Refinement (PMR) mechanism. These innovations enhance the suppression of complex noise. Experimental results demonstrate that the method enhances PESQ and STOI performance compared to the original GTCRN. It maintains a model parameter count of only 0.03 M. Thus, the method delivers strong speech enhancement and remains lightweight, making it suitable for real-time applications.