高阶MFCC的话者识别性能及其噪声鲁棒性

在一个以MFCC为特征参数的语音识别系统中,人们通常采用低阶的MFCC系数作为语音帧的特征矢量.本文对MFCC的高、低阶系数在与文本有关的话者识别中体现出的识别性能和噪声鲁棒性分别进行了实验分析,发现高阶的MFCC系数在干净环境下对于话者识别而言具有与低阶MFCC系数相当的识别性能,并且当环境信噪比恶劣时,高阶的MFCC系数表现出比低阶MFCC系数更强的噪声鲁棒性.基于这个结果,本文将高阶系数的取值范围进一步向低阶拓展,只滤除最易受噪声影响的几个系数,并与Delta参数相结合形成新的特征矢量.实验证明,这种经过适当选取的MFCC系数同时具有良好的话者识别性能和噪声鲁棒性.

The Recognition Performance and Robustness of High Order MFCC for Speaker Recognition

The low-order Mel Frequency Cepstral Coefficients (MFCC) is usually used in a speech recognition system when MFCC is selected to be the feature coefficient In this article, experiments are performed respectively on the performance and robustness of the high and low-order MFCC with a speaker recognition system. It is discovered that the high-order MFCC performs as well as the low order MFCC in clean environment ,further more, The high-order MFCC is more robust when the signal-to-noise ratio becomes lower, Based on this, the dimension of the high order MFCC is then expanded to 17, while the first 3 coefficients of MFCC are wiped off, and then all the 17 coefficients are combined with their delta MFCC to form the new feature vector. It is proved by experiments that this new feature vector not only performs well in clean environment, but also improves much on robustness in noisy environment.