基于方差归一化失真测度的改进的LBG算法

VQ Speaker Model Based on Robust Distortion Measure

矢量量化（ＶＱ）技术在话者识别系统中得到了广泛的应用。 ＶＱ码本的产生通常采用 ＬＢＧ算法，失真测度则为对矢量的各分量等权重的欧氏距离。在话者识别系统中特征矢量的各个分量的分布是有差别的，且对于不同的话者，这种差别的程度又是不一样的。由于不同分布的各维参数对话者识别的有效性各不相同，因此，文章提出了一种能反映这种有效性差别的失真测度，即：方差归一化失真测度。以该失真测度为基础，并结合时序相关的初始码本设计方法及有效的零胞腔处理技术，文章提出了改进的ＬＢＧ算法，同时利用该算法训练出改进的ＶＱ话者模型，并进行了话者识别实验。

VQ (Vector Quantization) technique is widely used in text-dependent and text-independent speaker recognition systems.The VQ codebook is usually generated by LBG algorithm. The distortion measure is Euclidean distance which gives the same weight to each coordinate. In this paper, a Robust distortion measure is presented. This distortion measure can reflect the different validness of each dimension with different distribution in feature space,so that it can not only improve the speaker recognition system performance but also enhance the system time robustness. We call it Normalized Square-Mean Distortion measure. Based on this distortion measure, an improved LBG algorithm is also presented in this paper. Better VQ speaker model can be generated using this improved algorithm.