복소 스펙트럼 마스킹 기반의 전장 환경에서의 잡음 및 잔향 제거

Abstract

Over the past decade, deep learning-based speech enhancement technologies have seen significant advancements and have found extensive applications in various environments. Recent developments have centered around deep learning techniques employing complex spectral masking, which consistently outperforms traditional time-domain masking methods. In this study, we harness complex spectral masking-based deep learning technology to remove noise and reverberation in a battlefield environments, making deep learningbased speech enhancement applicable in military operations and similar scenarios. Our proposed approach has exhibited superior performance across all three evaluation metrics, SI-SDR, PESQ, and STOI, when compared to conventional methods. Even when subjected to adverse conditions such as low SNR environments with nonstationary noise, our method co nsistently demonstrates impressive speech enhancement capabilities. Key Words : Speech enhancement, complex spectral masking, battlefield environments