Universal Score-based Speech Enhancement with High Content Preservation

Speech enhancement aims to improve the quality and intelligibility of speech signals corrupted by noise or distortions. Traditional methods often struggle to generalize across diverse noise types and acoustic conditions, limiting their real-world applicability. In this work, we propose a universal score-based speech enhancement framework that leverages recent advances in score-based generative modeling to robustly denoise speech signals while preserving critical speech content. Our approach models the complex speech distribution through a learned score function, enabling effective removal of various noise patterns without relying on explicit noise assumptions. Extensive experiments demonstrate that the proposed method achieves superior enhancement performance across multiple challenging noise scenarios, outperforming state-of-the-art baselines in both objective metrics and perceptual quality. Notably, the approach excels in preserving speech content and naturalness, making it suitable for practical applications such as telecommunication, hearing aids, and automatic speech recognition.