Listening to Speech in Mixture: Advances in Source Separation and Target Speech Extraction

Abstract

In this paper, the latest advancements in source separation and target speech extraction technologies are reviewed. The former technology separates individual sounds from an acoustic signal recorded with multiple voices and other sounds, and the latter one extracts only the speech of the desired speaker. These technologies make speech more understandable for humans and contribute to improving downstream speech applications. Two important approaches are discussed: signal-model-based and neural-network-based methods. Then detailed explanations of representative techniques in the approaches, blind source separation in reverberant environments, and target speech extraction based on voice features are provided. Finally, the future prospects of this technological field are discussed.