Enhancing long-distance speech signals using PDMA speech enhancement system

Abstract

In long-distance scenarios, the energy of speech signals rapidly attenuates with increasing propagation distance. Traditional acquisition devices alone find it challenging to capture weak speech signals. Even when acoustic amplification is used, the low signal-to-noise ratio (SNR) renders traditional speech enhancement methods ineffective. To address this challenge, this paper proposes a speech enhancement system that integrates both acquisition hardware and back-end algorithms. Specifically, the hardware module of the system consists of a parabolic reflector designed to amplify the sound pressure and a differential microphone array aimed at suppressing noise interference. Complementing this, the algorithmic component employs a deep-learning-based Generative Filter Summation Network (GfsNet), which effectively integrates signals captured from these two hardware modalities, thereby significantly improving speech quality. Simulation and comparative experimental results show that the proposed algorithm can achieve 8.43 dB SNR gain in part of the test for the presence of multiple noise interference sources in the long-distance scene from 10 meters to 80 meters. The Speech quality (PESQ) is improved to 3.12, and the intelligibility (STOI) is improved to 0.57 under the best conditions. The proposed system shows excellent enhancement results in complex and multiple noise source environments.