Discrimination Based On Support Vector Machine Speaker

Speaker recognition is also called "voice print recognition", belonging to a kind of biometric identification. As well as fingerprint recognition, iris recognition and face recognition, speaker recognition is used as a biological identification. Because of its universality, uniqueness, stability, accuracy and can be collected, etc, biological identification is often treated as one of the body’s biological characteristics significantly. Compared with other biological characteristics, voice, more representative, does not only contain these properties, but also produces at any time conveniently and can be collected through the microphone without special input devices when training and recognition. Thus it becomes a kind of biological characteristics people are willing to accept.Speaker recognition includes speaker identification and speaker confirmation. Speaker recognition is a kind of many-to-one process in which the speaker is identified in limited collection of speakers from a passage of voice. The speaker recognition system performance is affected by two factors. One is voice feature parameters, the other is the speaker’s recognition system. With the increasing scale of testers, the speaker recognition system performance degrades. The performance of voice recognition systems are directly affected by the speaker feature parameters selection. It has been proved that speaker recognition by SVM. The parameters selection of SVM plays a key role in the performance of recognition system. So extracting representative feature parameters with strong ability to identify and of small dimension, selecting reasonable parameters of SVM, can play a key role in the speaker recognition system performance.The research of following aspects is carried out in this paper:(1) For SVM classification method, comparing the recognition rate of the linear cepstrum parameters of LPCC, MFCC and the difference frequency MFCC, results show that system recognition rate is higher when selecting16-order MFCC as the characteristic parameters of speaker distinguishing. The first order difference combination characteristic parameters can improve recognition rate, but it will affect the efficiency of recognition system.(2) For SVM classification of MFCC, it is discussed the effect of the first two dimensional on the recognition performance. Results show that just removing the first dimensional characteristic parameters (energy coefficient) can help improve system recognition rate.(3) Normalizing MFCC, comparing the dimension reduction effect and recognition performance of principal component analysis and average value method, the result shows that the average impact value method put forward in this paper can effectively select the useful feature parameters in MFCC and can reduce dimensionality and enhance the effect of the voice recognition rate.(4) Combining the K-cross validation method, the grid search method, genetic algorithm and particle swam optization for parameter optimization of kernel function paramerter of SVM and penalty factor, identifying multiple speakers, the result shows when combining K-cross validation with web search method for parameters optimization of SVM, the recognition system has advantages in voice processing with high dimension and big sample data.