The Research Of Feature Parameters Extraction For Speaker Recognition

Speaker recognition as one of the biometric authentication techniques is to recognize speaker’s identity by its voice waveforms. With the rapid development of network and information technology, speaker recognition has caught more and more people’s attention for its application prospect, and has become a hot spot in the research field of biological authentication technology today. On the condition that research the feature extraction algorithms of the current speaker recognition, this thesis proposed an improved feature extraction algorithm based on MFCC and an auditory feature extraction algorithm based on gammatone filter banks. Finally, the simulation experiments were conducted in the MATLAB platform. The mainly works in this thesis are as follows:1. At first, this thesis introduces the related knowledge of speaker recognition technology in detail, such as the working principle, system structure and performance evaluation criteria and so on. Then the feature parameters extraction of speaker recognition was studied, especially for the extraction algorithm of LPC, LPCC and MFCC. Besides, the speaker recognition system based on GMM was also studied.2. An improved feature extraction algorithm based on MFCC was proposed. This thesis analyzed the advantages and disadvantages for MFCC and IMFCC. For the MFCC extraction algorithm’s shortcoming that traditional Mel triangle filter banks’computational precision insufficient in the high-frequency region, and the IMFCC extraction algorithm’s shortcoming that inverse Mel triangle filter banks’computational precision insufficient in the low-frequency region, this thesis proposed an improved filter banks that combined Mel triangle filter banks’low-frequency region with inverse Mel triangle filter banks’high-frequency region. Then in the MFCC extraction algorithm, we use the proposed filter banks to replace traditional Mel triangle filter banks, and got an improved feature NewMFCC. In addition, we also combined NewMFCC with spectral centroid, and extracted their combination feature parameter. Finally, the simulation experiments were conducted on TIMIT speech database; its results prove the effectiveness of the improved algorithm.3. An auditory feature extraction algorithm based on gammatone filter banks was proposed. This thesis analysis the composition of human ear auditory system and its working principle, and then the gammatone filter banks and the non-linear characteristics of auditory system were studied. In the new auditory feature extraction algorithm, we use gammatone filter banks to replace the traditional Mel triangle filter banks, and use exponential compression that about with frequency to replace the fixed logarithm compression; besides, we also employ half raised-sine cepstrum raised technology and got an auditory cepstrum coefficient GFCC. Finally, we compared and analyzed the performance of GFCC by simulation experiments, the results show that:compared with LPCC and MFCC, GFCC has better recognition rate and noise robustness.