说话人特征提取算法的研究  

Research on Extraction Algorithm of Speaker's Features

作　　者：孙鹏[1] 戴天虹[1] 李野[1] 

机构地区：[1]东北林业大学机电工程学院,哈尔滨150040

出　　处：《森林工程》2013年第2期143-147,共5页Forest Engineering

基　　金：黑龙江省自然科学资金项目(F200920)

摘　　要：研究说话人识别中特征提取算法,已有的线性预测系数(LPC)与梅尔倒谱系数(MFCC)目前都有着广泛的应用,可以说二者各具优点,但同时也各有其弊端。线性预测系数的低鲁棒性使其不能在噪声环境下得到准确的特征参数,而梅尔倒谱系数是基于人耳特性的,能够很好的保持噪声环境下的鲁棒性,将二者结合,研究一种新的提取特征参数的方法。分析其特性,并取其一阶差分与其相组合,运用基于高斯混合模型GMM的说话人识别系统检验其识别效果,同时比较其与MFCC、LPC的优劣。Among the feature extraction algorithms for speaker recognition, the linear prediction coefficients （LPC） and the Mel cepstrum coefficient （MFCC） currently have a wide range of applications. Both of them have their own advantages, but also have some disadvantages. Linear prediction coefficient has low robustness in noisy environment so that it cannot get the accurate parameters, while the Mel cepstrum coefficient is based on human auditory characteristics and maintains good noise robustness. In this paper, a new feature parameter extraction method was proposed in consideration of the merits of the two methods. The characteristics of the method were analyzed and combined with the first order difference. The recognition results were tested based on the speaker r, ecognition system in Gauss mixture model （GMM）. The comparisons with MFCC and LPC were conducted. According to the experiment, to observe the experimental results ultimately concluded.

关 键 词：说话人识别 特征提取 算法 

分 类 号：S776.0[农业科学—森林工程]