基于鲁棒听觉特征的说话人识别

 为了提高噪声环境中说话人识别系统的性能,本文提出了一种鲁棒听觉特征提取的算法,并将其应用到说话人识别系统中.运用自适应压缩Gammachirp滤波器组模拟人耳耳蜗的听觉特性,对输入的语音信号进行频域子带滤波,将得到的对数子带能量作为听觉特征参数.分别运用离散余弦变换和核主成分分析方法,对提取的特征参数进行特征变换,降低特征参数的维数,提高特征参数的噪声鲁棒性和个性表现力.实验结果表明,将提取的新听觉特征参数应用到说话人识别系统中,新特征参数在鲁棒性和识别性能上均优于梅尔倒谱系数和基于Gammatone的听觉特征参数.

Speaker Recognition Based on Robust Auditory Feature

In order to improve the performance of speaker recognition system in noisy environment,this paper presents an auditory feature extraction algorithm.It used adaptive compression Gammachirp filter banks to simulate the auditory characteristics of human cochlea,and the input speech signal was sub-band filtered in the frequency-domain.After logarithmic transformation,it can get the logarithmic sub-band energy as the auditory feature parameter.It respectively used discrete cosine transform and kernel principal component analysis method to transform the auditory feature and get the two new auditory features,which not only can reduce the dimension of the feature parameters,but also can improve the robustness and personality expression of feature parameters.The experimental results show that speaker recognition system with the new auditory feature parameters can get the better results in the robustness and recognition performance than Mel cepstral coefficients and auditory feature parameters based on Gammatone filter banks.