Abstract: A new method of using pitch period in Text-independent Speaker Identification System is presented. When training, estimate a Gaussian probability density function for pitch period of every person in the training library. When testing, use probability density calculated in the pitch period probability model to weight the likelihood measure of VQ or GMM. Then use the new likelihood measure to identify the testing speech. The experimental results show that this method can improve the recognition rate greatly compared with VQ and GMM. When the amount of codebook is 8, and the testing speech is 8s, this method can improve the recognition rate 13% compared with VQ.

Key words: vector quantization (VQ), Gaussian mixed model (GMM), raised-sine function windows, Gaussian function estimation for pitch period probability density, weighted likelihood measure

摘要： 研究了与文本无关的说话人辨认系统中一种新的使用基音周期方法。在说话人辨认系统中将矢量量化（VQ）、高斯混合模型（GMM）分类器结合，使用升正弦窗函数加权的线性预测倒谱系数（LPCC）。在训练时为训练集中的每个说话人估计一个一维高斯形式的基音周期概率密度函数；在识别时，将测试语音中提取的基音周期在训练集说话人基音周期概率模型中得到的基音周期概率密度对VQ、GMM分类器的似然测度加权，形成新的似然测度。实验结果表明，使用新的似然测度进行与文本无关的说话人辨认比VQ、GMM分类器的辨认率有较大的提高，码字个数为8，测试时间为8s时，辨认率相对VQ提高约13%。

关键词: 矢量量化(VQ), 高斯混合模型(GMM), 升正弦窗函数, 基音周期概率密度的高斯函数估计, 加权的似然测度