Research On Speaker Recognition Based On Feature Combination

Speaker recognition is a biometric authentication technology. The research aims at how to extract the speaker's voice characteristics to identify the speaker automatically. At present, speaker recognition has been applied in some fields due to its unique convenience, economy and accuracy, which has become a hotspot in speech processing. Feature extraction and pattern matching are the two keys of speaker recognition. This paper mainly focused on feature extraction method of the speaker recognition and the main work were as follows:(1) Based on Linear Prediction Coefficient(LPC) and Mel Frequency Cepstrum Coefficient(MFCC), Linear Prediction Mel Frequency Cepstrum Coefficient(LPMFCC) was proposed by integrating LPC into MFCC to realize speaker recognition, which don't increase the dimension of feature parameters and has Less computation. The LPMFCC can improve the recognition rate greatly.(2) In order to improve the accuracy of speaker recognition, multiple feature parameters can be adopted simultaneously. For the problem that each dimension comprehensive feature parameters have the different influence on the identification result, which to be treated equally is not necessarily optimal solution, a new method(FMLT parameters) to combine efficiently MFCC, LPMFCC and Teager Energy Operators Cepstrum Coefficient(TEOCC) based on Fisher criterion was proposed. FMLT parameters combines effectively the perception of the human ear,energy channel and nonlinear characteristics. And WFMLT parameters weighted by improved ascending half sine function based on FMLT parameters was put forward.Then the PCMLT parameters combine efficiently MFCC, LPMFCC and TEOCC based on PCA was proposed to further demonstrate the effectiveness of FMLT and WFMLT.(3) Aiming at the problem of the low calculation accuracy of MFCC parameters in high frequency, Inverted Mel Frequency Cepstrum Coefficient(IMFCC) that filters mainly in the high frequency part and Mid-frequency Mel Frequency Cepstrum Coefficient(Mid MFCC) that filters mainly in intermediate frequency part were adopted. Then this paper put forward the combination characteristic parameter(FMFCC parameters) based on Fisher criterion to combine MFCC, IMFCC and Mid MFCC to improve the performance of the speaker recognition system.(4) Study the speaker recognition system based on Gaussian MixtureModel(GMM) and Back-Propagation Neural Network, and apply the above feature extraction methods into the system to verify the effectiveness and feasibility of the combined feature parameters and system identification performance. Simulation results show that the recognition rates of FMLT compared with MFCC, LPMFCC,MFCC+LPMFCC, FMFCC and PCMLT is increased by 21.65%, 18.39%, 15.61%,15.01% and 22.70% in the pure voice database, and by 15.15%, 10.81%, 8.69%,7.64% and 17.76% in 30 d B noise environments. And the weighted WFMLT method has better recognition performance than FMLT, which enhances 2.62% in pure speech.The results show that FMLT and WFMLT can improve the recognition rate effectively and has better robustness.