Research On Speech Enhancement Algorithm And Its Application In Colored Noise

Speech enhancement is a technique which is that the original speech signal is extracted from noisy speech. It is widely used in many engineering fields. In practice, the disturbing noise is usually colored. Currently, it is necessary to further study and improve the enhancement of speech corrupted by colored noise.By studying classical speech enhancement algorithms such as Wiener filtering, spectral subtraction, signal subspace and Kalman filter in this thesis, it is found that they have better recovery results in the case of white noise. While they can’t effectively reduce the colored noise because the colored noise is non-stationary statistically. An effective pre-whitening technique is proposed to improve the de-noising performance of speech enhancement algorithms. It can deal with noisy speech more effectively in colored noise environment. The main work of this paper is to propose an improved subspace method and improved Kalman filter algorithm. In the improved subspace method, a whiting matrix based on the autoregressive(AR) process of colored noise is constructed,which changes the colored noise into the white noise. Then the subspace method in white noise is used to estimate the pre-whitened speech. The original speech can be obtained by the inverse of the whitening matrix. The pre-whitening technique overcomes difficulty in estimating covariance matrix of colored noise. In the improved Kalman filter, we also use pre-whitening technique to enhance speech corrupted by colored noise, which is multiplied by the whitening matrix. Then the whitened noisy speech is obtained. The standard Kalman filter algorithm is used to restore the speech. Different from the existing Kalman filter,the Kalman state transition matrix is not augmented. So it costs less computation. And an estimation of driving variance is given.In a word, two improved methods and their theoretical analysis are presented in the thesis. And we make a contrast with hot algorithms to confirm the superiority of ours in improving the quality of the speech signal by further experiments. In addition, the improved algorithms are successfully applied to the signal time delay estimation The time delay is estimated better after de-noising than before and more accurately than the existing enhancement algorithm.