Speech Bandwidth Extension (BWE) reconstructs missing high-frequency components in narrowband signals to enhance perceptual quality. Conventional signal-processing methods face performance limitations, while deep-learning approaches require substantial computational resources. To address these issues, we present HWB-Net, a novel Hybrid Waveform Bandwidth network that combines half-wave rectified signals with raw waveforms through a lightweight architecture, with a lightweight architecture, reducing both complexity and cost. Specifically, integrating half-wave rectification into BAE-NET-Lite enhances the model’s understanding of speech signals, leading to improved subjective and objective results, and replacing the decoder with a simple linear layer plus a weighted Gaussian mixture model (WGMM) significantly cuts down parameters (0.20M) and computational complexity (0.013G/s). Evaluations demonstrate HWB-Net’s competitive performance and practical viability for real-time communication through balanced efficiency-accuracy tradeoffs.