Study On Speech Enhancement Based On The RBF Networks

In general, speech signals are inevitably corrupted by various noises. These noises degrade the quality and the intelligibility of speech signals, seriously the processing systems couldn't work well. In order to minimize the effects of the noise on the performance of the processing systems, speech enhancement technology is applied in the various speech processing systems. Consequently the study of speech enhancement technology is very significant.This thesis discusses the speech enhancement technologies based on Radial Basis Function (RBF)networks, and focuses on the technologies based on double RBF networks in the frequency-domain. The fundamental and the implementation of the method and their improved forms are presented. Following is the main work of this thesis:1. By exploring the traditional methods, a new speech enhancement method based on RBF networks in the time-domain is proposed. This method can reduce the burden of the RBF networks and the training time efficiently. Simulation of the algorithm based on Matlab software is implemented. The results of the simulation prove that proposed method can effectively restrain noise and increase signal-noise rate (SNR). The experiment results indicate that the method can greatly improve the quality and the intelligibility of noisy speech, and have other advantages such as the widely applicable SNR range, less computation load.2. In the frequency-domain, double RBF networks are used to cut off the ingredient of noise. The first is used to train the formant coefficients and the second is used to train LPC coefficients. Then the modified spectrum envelope can be estimated by using these coefficients. At last the denoised signal can be reconstructed. The algorithm has it unique advantage. Particularly the method may maintain the preferable accurate of signal in speech waveform, and the speech is retained well, and the quality of speech signals have been improved obviously.3. Mel cepstrum distance (MCD) is suggested to evaluate the effect. Experiments show the method is more related with the intelligibility and outperforms the traditional SNR as it can offer more information regarding the applied conditions of enhancement approaches and their relative efficiency.