ISBN 978-3-8439-2918-9

978-3-8439-2918-9, Reihe Informationstechnik

Martin Krawczyk-Becker
Phase-Aware Single-Channel Speech Enhancement

174 Seiten, Dissertation Carl von Ossietzky Universität Oldenburg (2016), Hardcover, B5

Zusammenfassung / Abstract

Assistive listening devices and smartphones are often used in acoustically challenging situations, where the desired speech signal is degraded by acoustic noise. To reduce the undesired noise while maintaining the speech component, speech enhancement algorithms can be applied. Most common single microphone approaches work only on the spectral amplitude of the noisy signal, whereas the noisy spectral phase is effectively neglected and left unchanged. In this thesis, novel phase-aware algorithms are derived that overcome this limitation. Besides the enhancement of the spectral phase based on a new phase reconstruction method, clean speech estimators are proposed that explicitly utilize available phase information. Especially in challenging non-stationary noise scenarios, the additional phase information gives new means to distinguish desired speech components from noise outliers, resulting in an improved speech enhancement performance, which is confirmed by both instrumental measures and formal listening tests.