S³G-Net: Lightweight Banded Network for Real-Time Speech Enhancement

Real-time speech enhancement on resourceconstrained devices faces a persistent challenge: many causal systems rely on dual-branch magnitude/complex refinements or attention-heavy backbones, whereas ultra-light streaming models often lack explicit cross-band interaction and long-range temporal memory, resulting in unstable artifacts. We propose Streaming Subband State-space Graph-frequency Network (S³G-Net), a causal complex-masking architecture built around a subband state-space graph-frequency bottleneck. Specifically, S³G-Net (i) allocates frequency-dependent capacity through dynamic subband gating in the encoder with no look-ahead, (ii) restores cross-band structure via banded graph-frequency residual propagation that promotes locally coherent spectral evolution, and (iii) redesigns the temporal core into a lightweight hybrid architecture that integrates a dilated causal Temporal Convolutional Network (TCN) for long-horizon modeling with a graph-conditioned diagonal state-space module. The proposed graph-conditioned diagonal state-space model (SSM) uses per frame graph-frequency descriptors to modulate band-wise memory, enabling causal cross-band interaction with small attention while stabilizing streaming behavior. S³G-Net is implemented as a single-stream encoder-bottleneck-decoder with no auxiliary branches or post-filtering, and uses a model with only ∼30k parameters requiring ∼60M MAC/s. S³G-Net outperforms or matches substantially larger baselines in terms of perceptual quality across in-domain and out-of-domain evaluations, while maintaining competitive intelligibility at significantly lower cost.