| In real environment, the speech can be polluted by noise inevitably. Therefore,the quality of speech would decrease, and the intelligibility would drop. Therefore, inorder to use the speech efficiently, we should control the level of noise in speech.This thesis finished the following main works based on the Statistic model andBayesian theory:Firstly, we proposed a new speech amplitude estimator based on Bayesian theory.Aimed at the unprecise Gaussion model of speech and did the efficient improvementfor this by bringing in the super gaussion model of chi-square model to modeling thespeech amplitude before the derivation progress, and integrated with the newprobability density function as well as the beta-order perceptive error function in theprogress of deducing, then, we got a beta-order Bayesian estimator, which can exploitmuch more priori information of speech by simulating the speech amplitude moreaccurately. Experiment results showed that the new method can suppress the noisemore effectively, and gain great increase in segmental SNR(Signal to Noise Ratio),also, the quality can be ensured. The SNR improvement can be2dB in some situations,and above0.7dB commonly. From the spectrogram we can see that the “musicalnoise†was suppressed more effectively. Informal listening test also showed that thenew method can get more acceptable speech.Secondly, we proposed an improved spectral over-subtraction method for speechenhancement based on MMSE noise PSD estimation. Traditional short-time spectralover-subtraction speech enhancement methods are non-effective mainly due to thedifficulties in tracking the non-stationary noise, which leads to speech distortion ornoise residual, and degrade speech quality immensely. In order to describing the noisefeatures and gaining precise noise estimation, we proposed an improvedover-subtraction method. The new method combined the Power Spectral Density(PSD)estimation of noise based on minimum mean square error(MMSE) estimation theoryby using statistic model and Bayesian theory. Because most speech enhancementalgorithms heavily depend on the value of noise power spectral density (PSD), whichdecides the performance of speech enhancement algorithms. The proposed method cantrack both stationary and non-stationary noise well and deal with the residual noiseeffectively. The results showed that the proposed method can suppress the noise more accurately and gain good result for estimating clean speech. Otherwise, because of thelow computation load, it is very appropriate for real-time application such as DSP andother Embedded situation. Both segmental SNR and PESQ results showed that theproposed method gained better effects. The Spectrogram also showed the better resultin time and frequency domain. |
