Research On Speech Enhancement Algorithm Based On Noise Estimation

The single channel speech enhancement is a long standing problem in speech signal processing research. Now, the most effective single channel speech enhancement methods are based on the waveform processing method, i.e. the clean speech signal is estimated based on the filtering method. In the single channel speech enhancement, we lack the knowledge of the background noise, so the existing single channel speech enhancement methods can not meet the requirements of the practical application. The main problems are that the estimation of the noise is inaccurate, the enhanced speech has the distortion and there is lack of the effective noise suppression where the speech is absent. To solve these problems, this paper studies the speech enhancement method based on the estimation of the noise. Our research goal is to reduce the delay and the deviation of the estimation of the noise, to reduce the distortion of the enhanced speech, and to improve the noise suppression ability of the speech enhancement method where the speech is absent. We had accomplished innovative achievements as follows:(1) This thesis expounded the basic principle of the speech enhancement method and the estimate noise method based on short-time spectrum. The simulation results show that the enhanced speech by the spectral subtraction has a large number of "musical noise". The MMSE-STSA method effectively reduces the "musical noise" in the enhanced speech, but there are lots of the residual noise where the speech is absent. The MS has a lot of delay. The MMSE-SPP reduces the delay of the estimation of the noise, but its noise tracking ability remains to be improved.(2) We proposed a new method based on the minimum mean square error (MMSE) and Bayesian risk to estimate the noise. The method firstly deduced the estimate noise model using MMSE principle. The noise estimation model includes the linear estimation and the compensation of the estimation of the noise. We used the Wiener filtering method to calculate the linear estimation of the noise. The compensation of the noise estimation is determined by the low compensation quantity or the high cancellation quantity. Then, we used the Bayesian risk generated by the linear estimation of the noise to distinguish the deviation type of the estimation of the noise linear estimation. Finally, we used the miscalculation conditional probability between speech presence and speech absence to calculate the low compensation quantity and the high cancellation quantity of the estimation of noise. Compared with the typical method in the experiment, we proposed the estimate noise method which effectively reduces the tracking delay and log error.(3) We proposed a new speech enhancement method based on the maximum posterior probability and soft-decision using the super-Gaussian model. Firstly the method used the super-Gaussian model as the prior distribution of the actual speech signal. We derived the speech estimator using the maximum a posteriori probability when the speech is continuous. Then we calculated the speech presence probability of the current noisy frame by a soft-decision method. Finally we realized the speech enhancement when the speech signal is not continuous using the estimated noise. Compared with the typical method in the experiment, we proposed the speech enhancement method that reduces the speech distortion, and effectively suppresses noise where the speech is absent.