基于GMM_UBM的说话人确认系统的研究.pdfVIP

第 32 卷第 2 期 辽宁工业大学学报(自然科学版) Vol.32, No.2 2012 年 4 月 Journal of Liaoning University of Technology(Natural Science Edition) Apr. 2012 本刊核心层次论文 基于 GMM-UBM 的说话人确认系统的研究 霍春宝，张彩娟，赵红敏 （辽宁工业大学 电气工程学院，辽宁 锦州 121001） 摘 要：在基于 GMM 的说话人确认系统中，模型的训练是为每个说话人的语音建立模型，然后通过一定的 算法找到一组参数 ? ，使似然概率最大。文中通过对 GMM 的研究提出一种改进的模糊 C 均值算法(FCM)并将改 进后的算法应用到模型初始化中。同时，GMM 在话者确认时，语音数据不足会导致识别率下降，本文采用能覆 盖话者语音的高斯混合模型-通用背景模型(GMM-UBM)作为识别模型，通过算法比较及实验分析可知，改进算 法后的系统在识别率上明显优于传统的基于 GMM 的说话人识别系统。 关键词：说话人识别; 高斯混合模型; EM 算法; GMM-UBM 中图分类号：TP274 文献标识码：A 文章编号：1674-3261(2012)02-0098-04 Research on Speaker Identifying System Based on GMM-UBM HUO Chun-bao, ZHANG Cai-juan, ZHAO Hong-min （Electrical Engineering College, Liaoning University of Technology, Jinzhou 121001, China） Key words: speaker recognition; GMM; EM; GMM-UBM Abstract: In speaker recognition system based on GMM, model training was to build probability model for each speaker, then a set of parameters λ was found with a certain algorithm and which made likelihood probability maximum. After the GMM was studied, the improved Fuzzy C Means algorithm (FCM) was proposed and introduced to model initialization. In GMM speaker recognition, speech data deficiency can lead to recognition rate to be lower, so the GMM-UBM was used as the recognition model. Through algorithms comparasion and test analyses, the experimental results show that the recognition rate of the improved algorithm is obviousl