An improved β-order WEDM spectral amplitude estimator for speech enhancement

Abstract

This paper proposes an improved β-order Weighted Euclidean Distortion Measure (I-β-WEDM) spectral estimator for speech enhancement. The new β-order WEDM gain function is introduced, in which a novel method of computing order β is proposed. In this method, the input noisy signal is divided into several critical bands in the human auditory model. The value of order β is updated adaptively according to the signal-to-noise ratios (SNR) which are computed in different critical bands. The performance of the proposed estimator with different values of β has been evaluated by ITU-T G.160 under the white noise and the non-stationary noise, respectively. The experimental results show that, comparing with the reference algorithms, the proposed algorithm with adaptive β could provide large amount of noise reduction, and has a little impact on the level of speech signal. The quality of enhanced speech is improved at the same time.