Efficient β-order Perceptually Motivated Spectral Amplitude Bayesian Estimator Based On Chi-distribution for Speech Enhancement

Abstract

The traditional Bayesian estimator of short-time spectral amplitude is based on the minimization of the squared-error cost function under the common Gaussian probability density function (pdf). The Gaussian distribution, however, is not the optimal probability distribution. To overcome this phenomenon, we considered to replace the traditional distribution hypothesis of spectral amplitude of speech in this paper. More precisely, we proposed a β-order perceptive Bayesian spectral amplitude estimator which incorporated the assumption of Super-Gaussian chi-distributed spectral amplitude. The new weighting function incorporated the perceptive property as well as the different importance of the spectral valley and peak. Experiments showed that the proposed estimator can achieve a more significant noise reduction and yield a better spectral estimation over the most of latest enhancement algorithms.