Single Channel Phase-Aware Signal Processing in Speech Communication

Chapter 4Phase Processing for Single-Channel Speech Enhancement

Johannes Stahl and Pejman Mowlaee

Graz University of Technology, Graz, Austria

4.1 Introduction and Chapter Organization

The previous chapters have given an introduction on how to tackle the phase estimation problem in general and for speech processing in particular. In this chapter we will consider the knowledge gained in the context of speech enhancement, which has been an active field of research for decades. There exist well-established solutions to the problem of enhancing noise-corrupted speech, a wide range of them formulated in the STFT domain, motivated by its mathematical convenience and efficient implementation. The estimation of the clean speech STFT representation has been focused solely on processing the spectral amplitude while leaving the spectral phase untouched. From the perceptual point of view, an accurate estimate of the clean speech spectral amplitude is indispensable. In addition, the spectral phase was found to be perceptually irrelevant by Wang and Lim (1982) (see also Experiment 1.1 in Chapter 1), leading the subsequent research to neglect the information carried by the spectral phase. It is interesting that, at the same time, they state that a reasonable phase estimate could be beneficial if used for refinement of the amplitude estimation. More recent findings (Paliwal et al. 2011) revealed contradictory results, conceding perceptual importance to the spectral phase. Following this study, ...