Speech Enhancement Algorithm Research Base On Spectrum Attenuation

With the development of pattern recognition, artificial intelligence and other relevant technology, the requirement of the effect of speech signal processing and speech quality is increasing. But in real applications, the speech signal is inevitably corrupted by multifarious noises which degrade the performance of speech signal processing system seriously. Speech enhancement is a kind of specialized technology which is used for restoring the original voice from the noisy speech signal as undistortedly as possible.How to reduce the noise efficiently is a hard task in the domain of speech enhancement. Speech enhancement algorithm based on spectral attenuation is an excellent method, and it not only has high performance but also is easy to be implemented. Many speech enhancement methods have achieved good results in this framework. However, the promblems existed in spectral attenuation algorithm, including the accurate estimation about a priori SNR and voice distortion, still need to be done further studied. Aiming at these problems, the main works carried out in this paper are as follows:(1) A spectral subtraction based on priori SNR is proposed on the basic of analyzing the reason of producing "musical noise" using the traditional spectral subtraction speech enhancement methods. Firstly, a priori SNR is gained using a priori SNR estimation algorithm, then, according to the relationship between a priori SNR and a posteriori SNR, the gain function based on priori SNR is simply obtained using the priori SNR to replace the posteriori SNR of the gain function of traditional spectral subtraction. The method can not only suppress "musical noise" efficiently but also improve intelligibility of speech when processing noisy speech.(2) MMSE (Minimum Mean Square Error) is proposed to estimate a priori SNR to enhance speech. Based on the foundation of speech model and noise model, the priori SNR is estimated directly from noisy speech using MMSE. Performance is evaluated in terms of SNR, Itakura-Saito distortion measure, Perceptual Evaluation of Speech Quality, signal wave and spectrogram using the Weiner Filter speech enhancement method. Experimental results show that the proposed method is not only better than DD in the area of suppressing "musical noise" and preventing the voice distortion but also superior to non-causal a priori SNR estimation on real-time. The simulation results turn out that this method is effective.(3) In order to solve the problem that the enhanced speech has distortion caused by the loss of speech harmonic signal, an adaptive harmonic regeneration speech enhancement method is proposed. By doing a nonlinear transform to the enhanced speech, a harmonic retrieval signal is obtained. Combined with adaptive factor which is relevant to SNR, The signals including a harmonic regeneration speech signal adjusted by enhanced speech and the retrieval signal are obtained. The method overcomes the defect of the traditional harmonic regeneration method which uses a fixed adjustment factors has little recovery when the SNR is lower. The proposed method can effectively reduce the loss of the harmonic components and restrain speech distortion by the observation of spectrogram and the test of intelligibility.