Robust Speech Feature Extraction Reserch

Speech recognition is a means of human-computer interaction,in today's era of electronic computers so universal,how to allow the computer to accept human verbal instruction is a very challenging and meaningful subject.As with other machine learning problems in speech recognition,feature selection is the most important part of the start,has a distinguishing feature and stability is a prerequisite for a classification problem better recognition rate In order to get the characteristic of speech signal,the following work is done in this paper:1 based on the dimension reduction and normalization algorithm for non-specific human differences,a new algorithm is proposed.Effect of speech recognition rate is the biggest problem is non specific did not match the feature and background noise,for non specific differences in the past are normalized spectrum by VTLn,this paper proposed a combination of glottis,and channel characteristics of nonlinear return to a process of non structured by differences in the spectrum of a particular person and achieved better VTLN normalization effect compared,and combined with the idea of dimension reduction proposed a multidimensional differential improvement the performance characteristics of specific merging algorithm better.2 an improved supervised LPP algorithm is proposed for the high dimension nonlinear and class division of speech signals.Due to the temporal nature of the linguistic signal distribution,it is necessary to increase the distinction between basic elements.The traditional linear feature transform can not distinguish the nonlinear manifold structure of the different states of the same basic element,and the nonlinear transform can not meet the requirements of real-time performance.Therefore,this paper improves the distribution characteristics of the speech signal by using the linear transformation to approximate the distribution of the nonlinear manifold.3 the use of CNN neural network for the two time on the extraction of speech feature and recognition based on GMM-HMM model.The convolutional neural network(CNN)to solve the problem from the distinction between class scatter and the characteristic of speech signal and noise caused by the non specific person within class In the experiment of feature extraction using convolutional neural network,try to combine the mixed Gauss-Hidden Markov model and achieved better recognition results.