Model-based Techniques and Diffusion Models for Speech Dereverberation

Abstract

Reverberation degrades speech quality, especially for hearing-impaired listeners. Therefore, most speech communication systems (video-conferencing, smart home devices, hearing aids, etc.) now include dereverberation algorithms to increase the quality and intelligibility of speech. Traditional statistics-based dereverberation methods struggle in adverse conditions. In comparison, deep learning approaches show stronger performance, however they often lack interpretability and their failure cases are hard to predict. This thesis first explores hybrid models combining deep learning with domain knowledge -- called model-based techniques -- for optimal and robust dereverberation. The focus then shifts on the introduction of supervised diffusion-based generative systems in the design of dereverberation algorithms, while the last chapter unifies model-based algorithms and diffusion models for unsupervised dereverberation.