Noise suppression with high speech quality based on weighted noise estimation and MMSE STSA

Abstract

SUMMARYThis paper proposes a high speech quality noise suppression method based on weighted noise estimation and MMSE STSA. The proposed method continuously updates the noise estimate, using weighted noisy speech according to the estimated speech-to-noise ratio. In order to fully utilize the improvement offered by noise estimation, the spectral gain is corrected according to the estimated speech-to-noise ratio. By using accurate noise estimation, more accurate SNR than in the conventional method is obtained, which helps to reduce distortion in the enhanced speech. In subjective speech quality evaluations, the five-stage MOS was improved by 0.35 and 0.40 at the maximum, respectively, for the cases in which the speech was encoded and was not encoded after noise suppression. The improved version, which was developed on the basis of the proposed noise suppressor, satisfies all 3GPP minimum requirements for speech quality and has been installed in a commercially available model.

Analysis of MMSE Speech Estimation Impact in West Sumatra's Noises

“…Another estimation is obtained by weighted noise estimation (EDB) to implement the RSD estimating noise with noise on the heuristic weighting function [13]. The function of the optimal weight can also be obtained by measuring the average error minimum squares or minimum mean -squared error (MMSE) the [18], the MMSE method (ME) also has to consider aspects of bias based compensation noisy signal model whereas the compensation aspect bias heuristic way obtained on the method of [17].…”

A New Speech Enhancement Technique to Reduce Residual Noise Using Perceptual Constrained Spectral Weighted Factors

“…A priori SNR and noise power spectrum were estimated using the two -step a priori SNR estimator proposed in [15] and weighted noise estimation method proposed in [17],respectively.…”

Single Channel Speech Enhancement Techniques in Spectral Domain

“…Firstly, the noise variance was calculated by using the weighted noise estimator proposed in [19]. This method can update the estimated noise variance even if a speech signal exists.…”

“…Here, θ and γ H are constant values. The literature in [19] recommends that θ = 7 and γ H = 10. As shown in Figure 6, W(n) is a function of γ(n) given as…”