基于非线性幂函数的听觉特征提取算法研究

The Auditory Feature Extraction Algorithm Based on Power-law Nonlinearity Function

为提高说话人识别系统的识别率,提出采用非线性幂函数对人耳的听觉特性进行模拟,分别得到新的梅尔频率倒谱系数MFCC及其差分、加权倒谱系数.对得到的新的特征值进行增减分量分析,以获得高贡献值的倒谱分量,组成新的混合参数,使用高斯混合模型(GMM)分别对纯语音和三种典型噪声背景下的语音进行说话人识别,与传统MFCC相比,采用非线性幂函数改进的MFCC在识别率及鲁棒性上均有明显提高.

In order to improve the speaker recognition accuracy,the auditory characteristics of human are simulated by the power-law nonlinear function,and the new Mel frequency cepsral coefficients(MFCC)and its difference,weighted cepstral coefficients are obtained.The new characteristic values are analized from two angels that are increasing components and decreasing components,the vector with high contribution is drawn from it and new hybrid parameters are composed of them.GMM is used to recognize the speakers in four kinds of conditions which are pure speech and three kinds of typical noise background.Compared with the traditional of MFCC,New MFCC has improved the recognition rote and robustress.