Phase Aware Speech Enhancement and Dereverberation

Abstract:

This thesis aims to investigate the phase estimation problem in speech enhancement and speech dereverberation. Speech enhancement and dereverberation are critical in speech processing when dealing with speech with noise and interference. In many approaches to speech enhancement and dereverberation, only the magnitude is considered for enhancement, and noisy phase is used when reconstructing speech from enhanced magnitude. However, using noisy phase that is not consistent with the magnitude for speech reconstruction may degrade the quality of the re-synthesized speech, so this thesis focused on phase-aware methods. Firstly, we investigate issues of time alignment in the Perceptual Evaluation of Speech Quality (PESQ) measure. We demonstrate problems in the time alignment procedure in the PESQ measure when evaluating reverberant speech, and proposed a solution to overcome this drawback. We then propose a weighted magnitude-phase loss function for speech dereverberation to balance the estimation of magnitude and phase. It showed that the weighted magnitude-phase loss function improved the performance in both speech denoising and dereverberation when a suitable weight is chosen.  Next, we proposed a masked loss function for phase processing. With this loss function, we proposed to estimated phase as well as derivative phase features, and a phase reconstruction approach using derivative phase was also proposed. The reconstructed  phase was then used to re-synthesize speech with enhanced magnitude. The experiments showed that our proposed masked loss function improved the quality of re-synthesized speech.