Speech Enhancement Based On Time-domain Filter

Speech is usually contaminated by environmental noise in our daily lives. So many speech signal processing systems deteriorate sharply. Speech enhancement has been widely applied to other speech signal processing system as a front-end processing system. The objective of speech enhancement is to reduce the noise level and improve the quality of speech.Speech enhancement based on time-domain filter is a speech signal estimation method based on a time-domain error criterion. This method is one of the most widely used methods. It usually requires accurate parameter estimation. In this thesis, the following issues have been identified and explored:1. As decision-directed method only uses the data of the current frame and the data before current frame, it can not respond too fast to an abrupt increase in posteriori SNR, and it also can not smooth posteriori SNR enough in the low SNR area. The decision-directed method can not acquire the accurate priori SNR estimation in wiener speech enhancement method. In order to overcome the shortcomings of the decision-directed method, I propose a new approach for enhancement of noisy speech based a noncausal a priori SNR estimator. By using some data after current frame, the noncausal a priori SNR estimator can respond fast to an abrupt increase in posteriori SNR, and it also smooth posteriori SNR enough in the low SNR area. Through many simulations, it is demonstrated that the proposed approach can achieve better results.2. Kalman filter speech enhancement method is the method of tracking speech based on the minimum mean square error criteria. It uses speech generation model. For this speech enhancement method, the quality of speech enhancement bases on the estimation of model parameters. As kalman filter speech enhancement needs the linear prediction coefficients of clean speech, we need to carry out the pretreatment of noisy speech. In spectral subtraction speech enhancement method, random peaks often appear in the spectrum. In order to overcome the shortcomings of spectral subtraction, I propose a new kalman filter speech enhancement method based on minimum mean-square error log-spectral amplitude estimator (MMSE-LSA). Experimental results show the proposed method outperforms the kalman speech enhancement method based on spectral subtraction.