Colloquium - Details
You will receive information about presentations in time if you subscribe to the newsletter of the Colloquium Communications Technology.
Master-Presentation: Investigations on Phase-Aware Speech Enhancement Using Deep Neural Networks
Hennig Konermann
Mittwoch, 10th April 2024
1:30 PM
IKS 4G | hybrid
Speech enhancement aims to improve speech quality and intelligibility by removing noise from noisy speech signals. Currently Machine Learning (ML) based speech enhancement has become mainstream and is used in hundreds of millions of devices. This is crucial in various applications, from telecommunications to hearing aids. Historically, the phase component was considered unimportant for this task when using the analysis-modification-synthesis approach. However, with the rise of ML and, in particular, Deep Neural Networks (DNNs), these technologies have become increasingly important in recent times. This thesis presents an in-depth study of phase-aware speech enhancement using DNNs, initially focusing on the theoretical benefits of integrating phase information into the speech enhancement process through oracle experiments. A significant emphasis of this work is on the recently proposed Consistent-Inconsistent Phase (CIP) approach, discussing its advantages and disadvantages to phase estimation. Traditional magnitude estimation, with and without additional phase information, serves as the baseline for comparison. It has been demonstrated that CIP offers theoretical advantages over pure phase estimation and could, in theory, perform equally well as magnitude estimation without additional phase information while adopting the noisy phase for synthesis. However, the practical implementation does not fully realize its theoretical potential when validating the theoretical results by replicating the experiments with state-of-the-art DNNs. Solely in the context of background noise removal, a combination of magnitude and CIP estimation proves clear superiority to other techniques evaluated in this study. The estimation of the CIP emerges as a viable alternative to direct estimation of the clean phase, especially in noise-dominated signals.
