Restoring Harmonics: Enhancing Speech Quality with Deep Mask and Harmonic Restoration Network

While speech enhancement models have achieved remarkable progress in noise suppression, critical limitations persist in speech quality improvement. Current methods tend to over-suppress low-energy speech components and suffer from harmonic structure degradation. To tackle these challenges, we propose a two-stage model, Deep Mask and Harmonic Restoration Network (DMHRN). The Deep Mask (DM) stage utilizes a Noise-Reduced Mask (NRM) to aggressively suppress noise. Subsequently, the Harmonic Restoration (HR) stage leverages estimated fundamental frequency (F0) contours to guide the reconstruction of missing harmonic components. Experimental results on the DNS3 dataset demonstrate that DMHRN effectively restores the complete speech structure with fine-grained detail. It significantly outperforms baseline models, with a 6.01% improvement in P.835 OVRL and a 3.67% gain in P.808, achieving a substantial enhancement in speech quality without compromising noise suppression.