SigRNet: A Deep Intelligent Speech Receiver Based on Regression Model
Ziliang Zhou, Shilian Zheng, Zhijin Zhao, Xiaoniu Yang · 2024
In the realm of wireless digital communication, the reception of signals has always been a crucial component. Traditional receivers typically comprise multiple modules to decode signals that have been impaired by the physical channel. However, with the evolution of artificial intelligence and deep learning technologies, intelligent receivers based on neural networks have increasingly become a focal point of research. In this paper, we introduce a novel intelligent receiver based on a deep learning regression model, termed SigRNet, designed to replace all modules in the traditional signal reception process and directly output the restored analog voice signals. SigRNet is primarily divided into two parts: Backbone and Reconstruction. In the Backbone, deep separable convolutions are employed, integrated with three distinct attention mechanisms to extract features highly correlated with the input signal. In the Reconstruction part, these features are restructured using residual convolutions. Simulation results indicate that SigRNet, with less data transmission, outperforms traditional hard decision-making across various channel environments. Concurrently, in additive noise scenarios, SigRNet further reduces the quantization error inherent in digital signals.