Speech enhancement based on the Wiener filter and wavelet entropy

Abstract

Because of constant noise estimations, the speech enhancement of standard Wiener filters is poor under varied noise environments. In the present study, we propose an improved Wiener filter method for speech enhancement based on wavelet entropy. Wavelet entropy (WE) point detection can discriminate between speech activity segments and noise segments. This discrimination provides a basis through which noise can be estimated and updated accurately, leading to accurate a priori signal-to-noise ratios obtained from updated noise estimations, reduction of residual musical noise, and enhancement of speech signals degraded by non-uniform noise. Spectrogram comparisons of enhanced speech signals between the proposed WE Wiener filter and a standard Wiener filter show that the former is better at suppressing non-uniform noise than the later. Their perceptual evaluation of speech quality measures also show that the WE Wiener filter yields better enhanced speech quality than the standard Wiener filter.