基于自适应补偿的文本无关说话人识别  

Text-independent Speaker Recognition Based on Self-adaptation Compensation Transformation

作　　者：陈妮[1] 盛利元[1] 肖小清[1] 袁益民[1] 

机构地区：[1]中南大学物理科学上亍技术学院,湖南长沙410083

出　　处：《计算机仿真》2008年第6期277-280,共4页Computer Simulation

摘　　要：基于高斯混合模型的文本无关说话人识别系统通常采用最大似然算法。在纯净语音环境下,基于这种算法的说话人识别系统具有较好的性能。当系统的训练环境和测试环境失配时,这种算法的误识率急剧上升。针对帧似然概率的统计特性,提出了一种新的非线性补偿方法——自适应得分补偿法。通过对帧似然概率归一化、帧均匀化和重新排序赋值等系列补偿措施,改善了原算法的识别性能。实验结果表明,新的补偿方法能够降低误识率,在开集中平均可达20%,闭集中平均可达50%。Text - independent speaker recognition system based on Gaussian mixture model （ GMM ） usually uses Maximum Likelihood （ML） algorithm. In clear speech environments,the system based on this algorithm has good performances. If the background noise leads a mismatch between the training environments and the testing environments, the miss recognition rate of this algorithm will rise rapidly. This paper presents a self- adaptation compensation transformation for the frame likelihood probability. After taking a series of compensation measures like normalization, uniformity, re - assignment and scheduling to frame likelihood probability, the performance of the algorithm is improved. The results of experiments show that the transformation can reduce miss recognition rate up to 20% in open set and 50% in closed set, compared to ML.

关 键 词：说话人识别 高斯混合模型 自适应得分补偿 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]