The Application Of Time-frequency Analysis Methods In The Voice Signal Noise Reduction

In order to better specify the voice signal in the time-frequency, the Gabor transfer is involved. And the EMD method is investigated to reduce the noise in the signal. The thesis is mainly covered by the following parts: Firstly, the background and the importance of the paper are introduced, specifying the voice in the time and frequency domain is to better understand the voice signal and achieve viewing the full properties of the signal. The next part in the paper is the specification and comparison of the Gabor transfer and other mathematical tools to analyze the voice signal. The Fourier transfer is the most basic approach to decompose the frequency components. Admittedly, the method is widely spread; however, the temporal property is neglected. The short time Fourier with Gauss window, namely, the Gabor transfer is introduced to show the temporal property corresponding to the ripple in the frequency.For the Gabor transfer, the parameters in the algorithm are very important, because the result is very dependent on the shape of the Gauss window. The parameters are decided by the comparison of the effect in the processing of a piece of voice signal. The Gabor transfer needs a pro-process phase to maximum the specification capability to process voice signal.To overcome this problem, the empirical mode decomposition(EMD) is used. The method is fully data-depended, namely adaptive. By this method, the signal can be decomposed into many intrinsic mode functions or modes.The ensemble empirical mode decomposition(EEMD) is used as the solution for the ‘mode mixing’. To overcome the ‘mode mixing’ problem, the EEMD which adds the whole temporal space with white Gaussian noise can tell the very similar modes.To solve the problem, the EEMD is used, the voice can be decomposed into several different levels of the signal, to tell which signals are relative to the real physical vibration versus noise. The power spectral density method is involved. The tone can be recognized in the plot. However, the signal may be dominated by the extremely low frequency, it probably makes it impossible to tell the actual parts which determines the information of the signal.For this, the reverse cumulative power spectral density is investigated to tell which frequency dominates the intrinsic mode function. And the intrinsic mode functions, those are not in the frequency range we concern.Lastly, we applied the method to the voice signal ‘Wu Li Xue Yuan’, and tell the character of this signal. And the result of the experiment verifies that the method that we raise is right. The real part of the Gabor transfer is achieved to represent the time-frequency property of the voice.