Study On Speech Enhancement Based On Wavelet Transform And Constrained Variance Spectrum Estimation

Speech is the vital tool for information spreading and feeling expression, it's also the important bridge for communication between human beings and working machine. However, the speech is inevitably corrupted by the surrounding noise in the real life and the quality of the communication of speech information is therefore affected. Speech enhancement is an effective method for speech quality improvement and it has great application in the field such as speech recognition, speech low speed coding and speech communication between human and machine. The theoretical analysis and experimental research on the speech enhancement algorithm based on the wavelet package transform and minimum statistics spectral estimation methods are performed in this paper and the main works are explained as follow:The characteristics of current algorithms of the least statistic noise spectrum estimation are analyzed and a novel method which can automatically track the minimum short time power spectrum of noised speech is proposed. The vital part of this algorithm is a constrained variance smoothing filter which can effectively restrain the short time spectrum variance to reduce the estimation error caused by the minima tracking. We propose two methods to estimate the smoothing parameter. In the first one, the smoothing parameter is derived using the past value of the smoothed spectrum and another one is derived using the statistics of the smoothed spectrum. Compared with the other ordinary estimation methods, it can track the rapid alteration of noise spectrum without the need to judge the speech emerging possibility.The error compensation algorithm is proposed for the noise estimation of the least statistics and some factors that affect estimation error are analyzed, the compensation value is obtained using the fitting technique.The perceptual characteristic of human ear in bark domain is simulated and the critical bands wavelet package decomposing structure is designed in this paper. For speech signal sampled with 8kHz, che critical band contained 17 subbands with centre frequency interval is 1 bark. Based on this structure and connected with speech signal's characteristic, we dispart the signal below 625Hz and between 3kHz and 4kHz and let their centre frequency interval is 0.5 bark and receive 24 subbands at last. On the basis of critical bands decomposion, a new algorithm is proposed for speech enhancement using node threshold wavelet packet transform method, and this method uses soft threshold deals wavelet transform coefficient with different threshold, and spectral entropy method is applied to estimate the node noise. Masking effect of human ear on voice is researched, the noise masking threshold value is obtained by the analytical computation of enhanced speech energy.The gain wavelet package speech enhancement algorithm is proposed in this paper which divides the bias of enhanced speech to pure speech into residual noise and speech distortion obviously. For the shortcomings of many classic speech enhancement algorithm which focus only on the raise of SNR but ignore the problem of speech distortion, this paper constrains the gaining factor by minimize the speech distortion under the condition of the residual noise is less than noise masking threshold and designs the cost function and then receives a perceptual gain factor. We propose to control the speech distortion to be smaller than the residual noise in the power of wavelet packet coefficients to avoid noise's over-attenuation under low SNR and then receives a lower bound of noise masking threshold which eliminate the speech distortion furtherly.