Study Of Speech Enhancement Based On Principal Component Analysis And Independence Component Analysis

At present, the speech recognition system under quiet environment has already made satisfying performance. But in the real environment, the noise in the background will lead to the sharp drop of recognition rate which affect the performance of recognition system notably. Especially when the background noise contains some voice of other people, the system hardly work well and some strange characters will appear in the result. Therefore the key point of improving the speech recognition system relies on the effective enhancement of the voice itself.Aiming at the different background interferences, this paper gives corresponding solutions and improved algorithms which presented in the following aspects:(1) solved the'music noise'and the other problems of the traditional spectral subtraction algorithm under the additive white noise by using the speech enhancement algorithm based on signal subspace and comparing with the spectral subtraction algorithm. This algorithm is a enhanced algorithm for single-channel input and the additive noise, which firstly using the K-L transform to decompose the vector space of the voice signal with noise under the disturbance of the white noise into signal subspace and noise subspace. Then it adjusts the component value according to the energy of the eigenvectors'voice and noise signal after every K-L transform. By using the endpoint detection, it checks the noise frame which is used afterwards to estimate the energy of the noise of the coming voice frame. The transformed component value is adjusted by the rules of the optimization of the frequency field restriction, and finally the enhanced voice signal is estimated by using the reverse K-L transform.(2) The independent component analysis (ICA) derived from blind source separation (BSS) technique is a new method for the management of the multi-dimension signals, which uses the observed signals to restore the output source signal according to the statistic character of the input signal under the condition of unknowing the parameters of the source signal and the transmitting channel. Because the voice signals with noise can be regarded as a mixture of independent noise and the voice signals, therefore it is possible to enhance the voice signal by separating noise and voice signal through ICA. However the ICA requires that the observed signals are no less than the independent sources, which makes it is impossible to ICA the single channel signals directly but to collect voice signals through more than 2 channels. To solve this problem, two signal channels are generated and received ICA basing on the frequency decreasing algorithm. However the method dealing the single channel voice signal also has its limitation: it's only capable of dealing with the mixture of 1 independent source and 1 noise source, in the case of multi-signal mixture, it turn out to be useless.(3)Under the condition that the background noise is someone's voice, the ICA technology still able to isolate the specific targeted voice while the traditional methods of voice enhancement turn out to be useless. The basic idea of it is to establish a target function for the multi-dimension signals observed under the principle of statistic independence, and decompose these signals into several independent components through optimization algorithm in order to enhance and analyze them. Comparing to the traditional correlation removing technology of the double ranking space, ICA could not only remove the 1 and 2 ranking correlation between each component value, but is also capable of discovering and removing the high ranking correlations, which ensures the independence of the output component values. But the sequences of the source signal (independent component value) isolated through the ICA will remain uncertain, which brings trouble to the following applications. This problem is solved by using the Speaker Verification System to locate the target voice. This system contains the speaker verification technology based on the vector quantification and the codebook designed by using the classic LBG algorithm. By setting the specific speaker's voice character as the comparing object, it acquired the targeted voice signal from the isolated component values, and achieved the speech recognition rate of 90% in the experiment.(4) If the length of the mix signal is too long, we have to split up the signal by the time-frame when we use ICA to do the signal processing. So we have to intercept several times to finish the whole processing. According the theory of ICA , the order of the independent component is unknown. If we use the speaker Verification System to identify the target speech, the work amounts beyond compute and it is very difficult to put it into use. The paper put up with the new methord of correlation measurement. Although this paper has studied on some problems and some key technologies in the speech enhancement system, it's far from enough and there are still many problems which are requested deeple research and solution. However, due to the widespread application prospect of the speech enhancement system, the theory and the practice research in this paper will be contributed to the future work..