Recognition Of Human Speech Parameter Extraction Method Of Speaking

Speaker recognition is a biometric identification, also known as the voiceprint recognition. Speaker recognition is provided with important research value and widely application prospect in the field of human identity recognition.Extracting feature parameters that reflect the personality in speaker recognition is a key problem which will influence the identification results.This dissertation studies on the speech feature parameters extracting as well as the improvement of the parameters used in speaker recognition.In this paper, the three kinds of basic characteristic parameters about speaker recognition are used to study on the parameters extraction:MFCC, LPCC and PLPCC. The three kinds parameters are applied to the speaker recognition system. In this paper, the GMM model is used in the speaker recognition system. The identification results based on these three kinds parameters is analyzed.Further,this paper the pitch that reflect the inherent parameters of the speaker is combined with MFCC,LPCC and PLPCC. The results of experiments show that the combination features can effectively improve the recognition rate.Also,under the noisy environment, especially when the SNR is low the recognition rate will be decreased. This paper presents two kinds of improved parameter extraction method:parameter extraction based on the pitch synchronization and parameters extraction based on the reconstruction of voiced speech spectral. Pitch synchronous is based mainly on the pitch period of voiced speech signal.The variable length of the windows is usd in the pretreatment of the speech. If the voiced speech is truncated by non cycle length, that will led the spectrum leakage.Extracted parameters by this way will make the parameters in high frequency section more robustness. Voiced speech spectrum reconstruction according to the harmonic characteristics of short-term speech spectrum, and even in the noisy environment the harmonic characteristics will not be changed obviously. For this,the spectrum reconstruction can be used before the parameters extraction.This way can make the reconstruction spectrum closed to the true spectrum.The experimental results show that the above two kinds of improvement parameters can improve the recognition rate of the speaker recognition system,especially in the low SNR, such as Odb white Gaussian noise,the rate will be incread by15%to20%.