The Extraction Methods Of Characteristic Parameters Of Speaker Based On EMD

With the development of technology, to identify the speaker by machines is paid great attention by people gradually. Speaker recognition is to study how to make the human voice as a personal symbol, and different people are identified through the machine use this symbol. As an important technology that provide convenience to people’s daily life, speaker recognition technology has become a hot field of study.In speaker recognition, the speaker can be characterized by the parameters is the key to the research. To identify the speaker accurately, we should extract the characteristic parameters for good performance of speaker. Mel frequency cepstrum coefficient(MFCC) is not only a parameter based on the human auditory characteristic, but also one of the important parameters to describe the speaker’s characteristics.Empirical Mode Decomposition(EMD) is a suitable method for speech signal processing.Some meaningful results are obtained by applying the EMD in speaker recognition. In this paper,method to extract the characteristic parameters of the speaker recognition is studied in detail.Based on the present research this paper made the main work as follows:(1)FFT is used to convert the speech signal from time domain to frequency domain in MFCC extraction method. However, the instantaneous change of signal can not be reflected by using the Fourier transform. According to the situation, this process is improved in this article.Wigner-Ville Distribution is a description of the signal energy with time-frequency, and it can accurately reflect the time-frequency structure of the signal. But Wigner-Ville Distribution trapped in cross-term interference by multi-component signals generated. Take advantage of Wigner-Ville Distribution, Fourier spectrum is being replaced by a new spectrum that combining Wigner-Ville spectrum and Fourier spectrum for extracting MFCC parameter. A method based on time-frequency analysis to extract MFCC parameter is proposed. Experiments show that this method overcomes the cross-term interference, and get a precise time-frequency structure.Compared with the traditional MFCC extraction method, the correct rate of speaker recognition system is improved.(2)Because of the non-stationary and nonlinear characteristics of speech signal, a method based on EMD for speech feature parameters extraction is proposed. First, use short-time technology to screen out the unvoiced speech. Second, the voice is decomposed into a series of intrinsic mode functions by applying Empirical Mode Decomposition method. Each IMF contains only part of the information of the speech signal, and different characteristic information carried by different IMF component. At last, these IMFs are weighted to get a new speech signal forfurther processing. The purpose is to get the effective characteristics of speaker, and to screen out the useless information that can not characterize speaker. Experiments show that in speaker recognition system, the new method provides a higher accuracy than the traditional method.(3)With the above methods, based on the combination of EMD and time-frequency analysis,a new parameter extraction method is proposed. Namely the EMD is used to get the IMFs, and then weighted the IMFs for MFCC extraction by using the method of combination of FFT and Wigner distribution. After the method is applied to speaker recognition system, experiments show that, the method can greatly improve the accuracy of speaker recognition system and provides a better robustness compare with the two methods that mentioned above.(4) ASCC describes the information signals in the middle band, so use the method of EMD and Hilbert transform, the speech signal frequency in the middle band(1500Hz-2500Hz) are screened out to extract the ASCC parameter. Then combined with MFCC parameters, a new mixed parameter extraction way is proposed in this paper which is achieved by MFCC and ASCC based on Hilbert-Huang Transform. The experiments show that when the hybrid parameters and the traditional MFCC parameters are applied to speaker recognition system, the correct rate of the former has certain ascend.