- Authors:
-
Thieling,
L.
,
Jax,
P.
- Book Title:
- Proceedings of European Signal Processing Conference (EUSIPCO)
- Publisher:
- EURASIP
- Pages:
-
p.p.
111-115
- Date:
-
Sep. 2023
- ISBN:
- 978-9-46459-360-0
- ISSN:
- 2076-1465
- DOI:
- 10.23919/EUSIPCO58844.2023.10289814
- Language:
- English
Abstract
Speech enhancement algorithms usually operate in the short-time Fourier transform (STFT) domain and only enhance the magnitude spectrum, while adopting the noisy phase for synthesis. This is because the phase has often been considered unimportant. However, recent findings have proven otherwise, leading to an improved enhancement by considering the phase either implicitly or explicitly. In this paper, we propose a phase-aware extension of our recently published two-stage speech enhancement approach. It comprises, among other improvements, an additional explicit phase estimation stage whose structure is inspired by the fundamental ideas of our work on phase reconstruction. Unlike most phase-aware approaches, we do not estimate the clean phase but propose a novel combined consistent-inconsistent phase (CIP). It corresponds to a silence-generating phase for the noise-dominated time-frequency (TF) parts and thus allows noise reduction without modifying the magnitude spectrum at all. We show that this new CIP can provide a significant performance improvement compared to the clean phase. Experimental results confirm the effectiveness of our proposed extensions, ultimately leading to improved speech quality (PESQ, DNSMOS) and speech distortion (segmental SNR).
Download
BibTeX