Speech enhancement using generalized weighted β-order spectral amplitude estimator

In this paper, a single-channel speech enhancement method based on generalized weighted beta-order spectral amplitude estimator is proposed. First, we derive a new kind of generalized weighted beta-order Bayesian spectral amplitude estimator, which takes full advantage of both the traditional perceptually weighted estimators and beta-order spectral amplitude estimators and can obtain flexible and effective gain function. Second, according to the masking properties of human auditory system, the adaptive estimation methods for the perceptually weighted order p is proposed, which is based on a criterion that inaudible noise may be masked rather than removed. Thereby, the distortion of enhanced speech is reduced. Third, based on the compressive nonlinearity of the cochlea, the spectral amplitude order beta can be interpreted as the compression rate of the spectral amplitude, and then the adaptive calculation method of parameter beta is proposed. In addition, due to one frame delay, the a priori SNR estimation of decision-directed method in speech activity periods is inaccurate. In order to overcome the drawback, we present a new a priori SNR estimation method by combining predicted estimation with decision-directed rule. The subjective and objective test results indicate that the proposed Bayesian spectral amplitude estimator combined with the proposed a priori SNR estimation method can achieve a more significant segmental SNR improvement, a lower log-spectral distortion and a better speech quality over the reference methods. (C) 2014 Elsevier B.V. All rights reserved.