Deep Internal Learning for Single-Channel Speech Dereverberation

作者：Yuxi Zhang, Emilie d'Olne, Vikas Tokala, Patrick A. Naylor · 年份：2025 · DOI：10.23919/eusipco63237.2025.11226681 · 被引用次数：1 · 研究领域：Speech and Audio Processing、Hearing Loss and Rehabilitation、Advanced Image Processing Techniques

Deep Internal Learning (DIL) is a paradigm with high potential in deep learning, as it reduces reliance on external training data and uses a lightweight model compared to traditional deep learning methods. Since its introduction, DIL has been explored in the field of image processing, including applications such as image super-resolution, denoising, and deblurring. However, its potential for other signal-processing tasks such as speech enhancement remains relatively underexplored. In this study, we investigate the feasibility of utilizing DIL for single-channel dereverberation. Specifically, we develop a small speech-specific Convolutional Neural Network (CNN) that is trained exclusively on example pairs derived from the observed reverberant speech itself. We evaluate our approach in the context of a dereverberation task in oracle experiments and more practical scenarios, considering rooms with a range of reverberation times. Experimental results demonstrate that the DIL approach can achieve higher speech enhancement and dereverberation scores compared to the traditional Weighted Prediction Error (WPE) algorithm. This paper showcases the potential of DIL for speech enhancement and formulates several open issues for future research on this topic.