The Research Of Speech Enhancement Algorithm

Nowadays, speech enhancement is an important technology in speech signal process fields, it is widely used in digital speech systems about speech discriminating, speech composing, speech coding. The purpose of speech enhancement is that the pure original speech is extracted from the noisy speech as much as possible. However, the noisy speech signal is generated randomly, which completely eliminated is almost impossible. Therefore, the actual goal of speech enhancement is as follows: to improve speech definition, so as to improve speech quality; to increase speech intelligibility, so as to facilitate listener understanding.For the moment, there are a lot of speech enhancement methods, for instance: based on short-term spectrum estimation, masking effect, signal subspace, wavlet packet and etc.This paper mainly elaborates the basic principle of the short-time spectral estimation speech enhancement methods, based on subtracting spectrum, MMSE (minimum mean squaree error), and wiener filter. Simulation results show that wiener filter method can more effectively depress the musical noise.This paper describes the basic principle of the signal subspace speech enhancement methods, based on subspace method of time domain constraint estimator(TDC) and frequency domain constraint estimator(SDC). A speech enhancement method based on the combination of the signal subspace and wiener is proposed, the wiener filter method is used in the signal subspace speech enhancement algorithm under the white noise. Simulation results show that under the background of white and train noise, the SNR in this method is more excellent than in the traditional subspace method, the musical noise is depressed effectively. In this paper a speech enhancement method based on the combination of the signal subspace and hearing masking effect is proposed, by the subspace filter the noisy speech is enhanced, by an auditory post-filter based on hearing masking properties of Johnston model, the enhancement speech which is consistented with human auditory characteristics is smoothed , the pure speech is gained. Simulation results show that under the background of white and train noise, the SNR in this method is more excellent than in the traditional subspace method, the musical noise is depressed effectively. Meanwhile the algorithm complexity is low, computation speed is fast, so it is beneficial real-time implementation.This paper elaborates the basic principle of wavelet packet transform speech enhancement method, in the method, by the 5-order Daubechies wavelet, the noisy speech is deal with Bark-scale wavelet packet decomposition, which is simulated with the human ear's auditory characteristics, mainwhile a new threshold function is used to improve the method. A speech enhancement method based on wavelet packet and hearing masking effect is proposed. In this method by an improved wavelet packet transform filter, the noisy speech is enhanced, by Johnston hearing masking model the hearing noise masking threshold is gained, by the perceptual filter based on hearing masking effect, the enhanced speech is smoothed, so the better clean speech is gained. Simulation results show that under the background of white noise, the SNR and PESQ in this method is more excellent than in wavelet packet method, the musical noise is depressed effectively. Mainwhile, under the case of low SNR, the enhancement effect is more notebal. This paper proposes a speech enhancement method based on signal subspace and wavelet packet transform. In subspace domain the noisy speech eigenvalue is filtered by the wavelet packet transform method, the clean speech eigenvalue after filtering is gained. The enhancement speech is gained by subspace inverse transform. Simulation results show that this method is more excellent than the traditional wavelet packet transform method, the algorithm complexity is low, computation speed is fast.