基于MFCC特征组合参数的说话人识别研究

Speaker Recognition Research Based on the Combination of MFCC Feature Parameters

为提高说话人识别系统的识别率,提出了一种提取Mel频率倒谱系数(MFCC)与差分特征组合参数的方法:先对传统的MFCC参数进行特征分量归一化处理,提升MFCC系数的噪声鲁棒性;再用高斯混合模型(GMM)构建了说话人识别系统。使用TIMIT语音库进行实验测试,并比较了不同高斯混合数的MFCC特征参数组合对识别率的影响。结果表明:使用改进的MFCC混合参数明显地提高了说话人的识别率。

To improve the recognition rate of speaker recognition system, this paper presents a method to extract the combination parameter of Mel frequency cepstral coefficients （MFCC） and accelerated coefficient. The normalized eigenvector of the traditional MFCC feature parameters, improved noise robustness of MFCC coefficient. The paper constructs a speaker recognition system using Gaussian mixture model （ GMM ） , and conducts on TIMIT corpus, and analyzs recognition ratios under different Gaussian mixed numbers （GMN） and the combination of MFCC feature parameters. Experiments show that feature parameters combined improved MFCC with accelerated coefficient are beneficial to improve the speaker recognition rate.