论文部分内容阅读
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。
Vector Quantization (VQ) technology has been widely used in speaker recognition system. VQ codebook generation usually LBG algorithm, the distortion measure is the weight of the vector components of the Euclidean distance. There is a difference in the distribution of the individual components of the feature vector in the speaker recognition system, and the degree of this difference is different for different speakers. Since the effectiveness of interrogators in different dimension parameters is different, the paper presents a measure of distortion that reflects this validity difference, namely, variance normalized distortion measure. Based on the distortion measure, combined with the design of initial codebook and effective zero cell cavity processing technique, an improved LBG algorithm is proposed and an improved VQ speaker model is trained by this algorithm. Speaker recognition experiment.