Research On Improved Algorithm Of Feature Parameter Extraction Based On Speaker Recognition

The extraction of feature parameters is the key part of speaker recognition. This dissertation focuses on the modified extraction algorithms of feature parameters of speech signal. We mainly do some research work on the modified extraction algorithms of Mel Frequency Cepstral Coefficient (MFCC) and Linear Predictive Cepstral Coefficient (LPCC). Then these modified parameters are tested on the GMM based speaker recognition system. The main research contents are as follows.In order to conquer the drawback of the conventional MFCC which loss high frequency information of speakers, a modified parameter based on Empirical Mode Decomposition (EMD) and self-adaptive high frequency weighted method is presented. As the ranges of different speakers’speech signals are not identical, it is hard to define the high frequency domain of speech signals. Thus how to partition the high frequency domain to be weighted is extremely needed. We propose a reasonable definition of high frequency domain and verify the first Intrinsic Mode Function (IMF) of speech signal contains the high frequency information defined by the proposed method. Then we fuse the speech signal with the first IMF by a weighted coefficient such that the high frequency information of speech signal can be enhanced. Finally, the modified MFCC is computed from the fused speech signal. The experimental results show that the recognition rate of the proposed MFCC based speaker recognition system is higher than that of conventional MFCC based one and also robust to the noisy speech signals.We also study the contribution of each dimension of cepstral parameters to speaker recognition system and employ Fisher Ratio method to compute the contribution of them. Then based on the contribution of each dimension of parameters, weighted method is used to improve the performance of the cepstral parameters. The results of experiments demonstrate our method can improve the performance of MFCC and LPCC for speaker recognition system and the modified MFCC (LPCC) computed by the proposed method based speaker recognition system are more robust.