基于VQMAP模型和AdaBoost学习算法的说话人识别  

Speaker recognition based on VQMAP model and AdaBoost learning algorithm

作　　者：吴海洋[1] 吕勇[1] 吴镇扬[1] 

机构地区：[1]东南大学信息科学与工程学院,南京210096

出　　处：《东南大学学报（自然科学版）》2010年第3期476-480,共5页Journal of Southeast University：Natural Science Edition

基　　金：国家自然科学基金资助项目(60971098)

摘　　要：为了解决传统说话人识别系统在集成学习后识别速度变慢且容易过学习的问题,构造了一种基于最大后验矢量量化(VQMAP)模型和自适应提升(AdaBoost)学习算法的说话人识别系统.首先,分析了说话人识别系统中基分类器性能对集成分类器泛化误差的影响.然后,针对说话人的类别数,构造适当精度的VQMAP模型.最后,利用包含提前终止策略的AdaBoost学习算法将该模型提升为强分类器.实验结果表明:该算法的识别速度较高,是最大后验高斯混合模型(GMMMAP)的9倍;该算法可有效控制AdaBoost学习算法在说话人识别中的过学习问题,其性能优于VQMAP模型,且在训练数据较少或者类别数可预计的情况下,其性能可接近甚至超过GMMMAP模型.In order to solve the problem of low recognition speed and overfitting resulting from ensemble learning in traditional speaker recognition systems,a novel speaker recognition system based on the maximum a posteriori vector quantization model（VQMAP） and the adaptive boosting（AdaBoost） learning algorithm is presented.Firstly,the influence of base classifier performance on the generation errors of the boosted classifier is analyzed in the speaker recognition system.Then,a suitable VQMAP classifier matching the speaker number is constructed.Finally,it is boosted to a strong classifier by the AdaBoost learning algorithm with the early stopping method.The experimental results show that the proposed algorithm has a faster recognition speed,which is 9 times faster than that of maximum a posteriori adapted Gaussian mixture model（GMMMAP）.It also reduces the overfitting of the AdaBoost learning algorithm in speaker recognition.The performance of the boosted VQMAP model is better than that of the VQMAP model,and in the case of limited data or a predictable speaker number,it can reach or exceed the GMMMAP model.

关 键 词：最大后验矢量量化模型 自适应提升 提前终止 说话人识别 

分 类 号：TN912.34[电子电信—通信与信息系统]