基于听觉模型的话者特征参数提取及其在噪声背景下的话者辨识

本文基于人耳听觉模型提出了一种鲁棒性的话者特征参数提取方法。该种方法中，首先由Gamma tone听觉滤波器组和Meddis内耳毛细胞发放模型获得表征听觉神经活动特性的听觉相关图。由听觉神经脉冲发放的锁相特性和双声抑制特性，我们将听觉相关图每个频带中的幅值最大频率分量作为表征当前频带特性的特征参量，于是所有频带的特征参量便构成了表征当前语音段特性的特征矢量；我们采用DCT交换进一步消除各个特征参量之间的相关性，压缩特征矢量的维数。有效性试验表明，该种特征矢量基本上反映了输入语音的谱包络特性；抗噪声性能实验表明，在高斯白噪声和汽车噪声干扰下，这种特征参数比LPCC和MFCC有较小的相对失真；基于矢量量化的文本无关话者辨识表明，对于三种类型的噪声干扰该种特征参数在低信噪比下都获得了较好的识别结果。

Speaker feature extraction based on human auditory model and speaker identification under noisy background

This paper proposes a robust speaker feature extracting algorithm based on human auditory model. In the algorithm, we first obtain the auditory correlogram from the Gamma tone filter bank and the Meddis inner hair cell model. Then according to the phase lock characteristic and two-tone inhibition phenomenon of the auditory nerve firing activity, we select the most dominant frequency component to characterize each frequency channel of the auditory correlogram, and get all these principal frequency com-ponents across channels for the feature vector of the current speech frame. To reduce correlation between the elements and dimensions of the feature vector, DCT transfor-mation is used. The feature effectiveness experiment shows that this feature represen ts the speech spectral contour basically, while the anti-noise experiment indicates that this feature has smaller relative distortion compared with LPCC and MFCC parameters under Gauss white noise or car noise interference. The speaker identification based on vector quantification shows that this speaker feature based system performs better than those LPCC and MFCC based system, especially for low signal to noise rate.