Speech enhancement by minimum mean-square error spectral amplitude estimation assuming weibull speech priors

Abstract

In this paper, we present a novel single-channel speech enhancement method in the Discrete Fourier Transform (DFT) domain. Here, the amplitude of DFT coefficients of a clean speech signal is modeled by a Weibull probability density function. Measuring the Jensen-Shannon divergence (JSD), Weibull distribution showed a better fit to clean speech signal compared to the previously fitted distributions such as gamma and Rayleigh. Therefore, we modify the Minimum Mean Square Error (MMSE) estimation algorithm for speech enhancement considering Weibull speech priors and Gaussian additive noise signals. The enhanced speech signals are assessed based on the perceptual evaluation of speech quality (PESQ) and segmental signal-to-noise ratio (SEG-SNR) criteria. Extensive simulation experiments on speech signals degraded by various additive non-stationary noise sources demonstrate that performance improvements are possible employing Weibull speech priors in the MMSE-based speech enhancement algorithm compared to the Rayleigh and Gamma PDFs.