Speaker Recognition Based On MFCC And IMFCC

Speaker recognition is a kind of technology that take the use of the information contained in the speech signal to identify who is the speaker or to confirm whether the speaker is the claimed one. Now in low-noise and low distortion environment, speaker recognition has performed quite perfect. But noise is everywhere in the practical application environment, which make the speaker recognition rate dramatically declined. As a result, extracting the robust feature and designing the effective classifier to get a good performance in the noise environment have become a hotspot, and then realizing the speaker recognition system come to practice from laboratory.In the view of the above requirements, this paper design a system in the background of shortwave channel, use the speech signal obtained from the shortwave-channel environment as the experimental data. In the feature extraction part, analyze MFCC which based on the human auditory mechanism, due to the structure of its filter bank, it captures characteristics information more effectively in the lower frequency regions than the higher regions. Thus, there must be information contained in the high frequency is missed. This work gets a new set of features by inverting the filter bank structure which can make up the lack of MFCC. Considering the complementary relationship of the two features, design a combining classifier decision system, which contains two sub-classifiers, uses MFCC and IMFCC as the features respectively, SVM as the classifier, then combine the decisions of the two sub-classifiers by using a function. Finally, achieves the aim to improve the performance of the speaker recognition system.In addition, presents an open-set speaker-recognition arithmetic based on speaker adaptive dynamic threshold. Make a model of a non-specific speaker R_N+1 which integrated all the reference speakers' characters, add it to the set of the reference speakers, whether the speaker is in or out the set depends on the score of R_n+1, if in give the result of the recognition, otherwise add it to the set of the reference speakers.The experiments show the method proposed get a better performance in the noise environment.