LOOK ATTENTIVELY TO HEAR: AUDIO-VISUAL SPEAKER EXTRACTION

Abstract

Human has the inherent ability to selectively listen to a speaker in the presence of interference speakers and background noise. Speaker extraction algorithm mimics such selective attention to extracting only the target speech signal in an adverse acoustic environment, which serves as an indispensable front-end to many speech applications. This thesis focuses on designing audio-visual speaker extraction algorithms to improve the quality of speech signals in real-world adverse conditions.