کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

GSE-Net: Multi-Scale Feature Extraction for Single-Channel Speech Enhancement

نویسندگان :

Alireza KayedNezami¹ Rahil Mahdian Toroghi² Hassan Zareian³

1- دانشگاه صداوسیما 2- دانشگاه صداوسیما 3- دانشگاه صداوسیما

کلمات کلیدی :

Speech enhancement،Single Channel،Transformer،Channel Attention،GPCFormer

چکیده :

Single-channel speech enhancement remains a central challenge in signal processing, where the goal is to recover clean speech from signals contaminated by environmental noise. Although CNN–Transformer hybrids have shown strong performance, the field still requires models capable of capturing complex time–frequency dependencies and extracting multi-scale representations. This paper introduces GSE-Net, a deep hybrid architecture that combines the local feature modeling capabilities of convolutional networks with the long-range dependency modeling of attention-based modules. This design strengthens the model’s capacity to jointly capture global and local structures. Experimental evaluations on the VoiceBank+DEMAND dataset show that GSE-Net delivers results that are comparable to, and in many cases surpass, leading state-of-the-art methods. The model maintains high speech intelligibility and significantly enhances perceptual quality, effectively advancing single-channel speech enhancement toward near-clean speech reconstruction.