Speech Enhancement Integrating Linear Prediction And Wiener Filtering Technique And Its Implementation

In the past few decades, many methods have been proposed for solving the problem of speech enhancement. Among these methods, the codebook driven Wiener filtering using linear prediction performs better than traditional ways in nonstationary noisy environment. However, the considerable residual noise in the enhanced speech obtained by codebook method is a problem. Therefore, finding an effective way to reduce the residual noise in enhanced speech with codebook driven method becomes a serious problem. In this paper, a relatively complete set of speech enhancement scheme is proposed by integrating the method of codebook driven Wiener filtering using linear prediction and the method of traditional SNR estimation based Wiener filtering.The research work of this paper contains the following three parts:Firstly, based on the existing codebook driven method, we proposed a method integrating linear prediction and Wiener filtering for speech enhancement. T he fact based the proposed method is that the codebook driven method could preserve speech components better in high SNR regions, while SNR estimation based Wiener filtering method performs better in low SNR regions. To implement the proposed method, the gain functions of the two methods are constructed respectively in the integrating method firstly, and then a sigmoid function is used to weight these two gain functions. The new constructed gain function reflects the characteristic of codebook method in high SNR regions and that of SNR estimation method in low SNR regions. By taking advantages of the two approaches and avoiding their disadvantages, the proposed integrating method performs better than both of them.Secondly, to solve the problem of the distortion of spectral shape in unvoiced frames and that individual speech drowned by noise, a harmonic enhancement method based on speech classification is proposed. This method uses the ways of speech energy analysis and pitch detection to classify the speech by frames. Then appropriate enhancing strategies are used according to different type of speech. Overall, the proposed method has a better subjective and objective quality than the integrating method.Finally, the integrating method is applied to the front end of G.718 codec. Testing results show that the proposed method performs better than the enhancement module in G.718 codec in noisy environment.