TSRH-Net: A serial time-frequency backbone with SNR-gated refinement for extremely low-SNR speech enhancement

In complex acoustic environments, low signal-to-noise ratio (SNR) conditions severely degrade speech structures and reduce speech intelligibility. These conditions also place high demands on the robustness and generalization performance of speech enhancement models in unseen noise environments. To address these challenges, we propose TSRH-Net, a novel network architecture tailored for speech enhancement under extremely low-SNR conditions. TSRH-Net improves speech feature recovery and stabilizes spectral reconstruction under severe noise interference by constructing a speech-structure-oriented serial time–frequency enhancement backbone and incorporating a residual recovery mechanism that combines feature-domain refinement with complex spectral compensation. Experimental evaluations on the DNS-Challenge dataset under multiple SNR distributions and on the NoiseX-92 cross-noise dataset show that TSRH-Net achieves consistent improvements in low-SNR speech enhancement tasks and demonstrates improved robustness to unseen noise conditions. Audio samples are available at https://basionwu.github.io/TSRH-Net/ .