The Research Of Speech Enhancement Algorithms Based On Spectral Estimation Statistical Model

In the real word, the speech signal is inevitably subjected to the interferencewhich comes from the ambient noise environment. To restore the original speechsignal from the noisy speech as much as possible, enhance the speech quality andimprove the speech intelligibility are the main goal of speech enhancement. There arenumerous kinds of speech enhancement algorithms, but the speech enhancementwhich is based on the spectral estimation statistical model is becoming one of themost widely used technologies. Because it is not only simple and easy to understand,but also applicable to a large range of SNR as well as wide noise environment.This paper studies the speech enhancement which is based on the spectralestimation statistical model, and proposes two novel speech enhancement methods:In view of the different performance of the minimum mean square error (MMSE)estimators of the magnitude-squared spectrum (MSS) and short-time power spectrum(SP), at high a priori SNR and low a posteriori SNR levels, the MMSEmagnitude-squared spectrum (MMSE-MSS) estimator markedly reduces the residualnoise without introducing speech distortion encountered in speech processed byMMSE estimator of short-time power spectrum (MMSE-SP). However, MMSE-MSSestimator can’t reach the better performance than the MMSE-SP estimator in otherSNR levels, therefore, a hybrid gain function based on the two estimators is proposed.Moreover, owing to the existence of the uncertainty of speech-presence in the noisyspeech model, hence, this paper estimates the priori speech absence probability ofeach frequency point and combines the resulting gain functions for furtherimprovement. Numerous simulation results demonstrate that the novel method cansignificantly suppress background noise and improve speech intelligibility.In Bayesian approaches for speech enhancement, minimizing the Bayes risk iscommonly used in the treatment. Different cost functions using Bayesian estimationcan derive different gain functions, therefore, various corresponding cost functionhave been proposed. However, the absolute error function has been paid less attention.In this paper, a novel conditional median magnitude-squared spectrum speechenhancement method which is motivated by the MMSE magnitude-squaredspectrum(MMSE-MSS) estimator is proposed. Compared with the MMSE-MSSestimator, this method gets better speech intelligibility and perceptual quality.