距离加权矢量量化文本无关的说话人识别

Text independent speaker recognition based on VQ and modified by the weighted distance

本文在研究说话人识别的矢量量化方法时，分析了用矢量量化建立说话人识别模型的可行性。针对量化码本描述的不完全性，提出了一种经距离加权的矢量量化方法，能更好地刻划出说话人语音特征空间的精细结构，从而提高正识率。本文还对特征参数ＬＰＣＣＥＰ的选取进行了理论分析和实验研究，提出了平均互—自差异比的概念，给出了一种对特征矢量的每一维分量识别能力进行定量化估算的公式。实验结果表明，距离加权矢量量化是一种具有很高正识率的与文本无关的说话人识别方法。

This paper analyses the feasibility of building the model based on vector quantization (VQ) in speaker recognition.We propose a weighted modification VQ to overcome the incompleteness of the description of the code book.By this method we can give a better description of the fine structure of the Speech Feature Space of the speaker and improve the recognition rate.We also theoretically discuss the selection of LPCCEP feature parameters and have verified it by experiments.Then we present a conception of the average mutuality self variance proportion and a formula for quantitatively estimating the recognition capability for each heft of feature vectors.The result of experiments indicates that the weighted modification VQ is a kind of highly efficient method with high performance for the Text |Independent Speaker Recognition.